- Email: [email protected]
Mechanical imaging:
International Journal of Medical Informatics 49 (1998) 195 – 216
Mechanical imaging: A new technology for medical diagnostics Armen Sarvazyan Artann Laboratories, 1 Ri6a A6enue, North Brunswick, NJ 08530 -3302, USA
Abstract Mechanical imaging (MI) is a newly developed modality of medical diagnostics based on reconstruction of tissue structure and viscoelastic properties using mechanical sensors. The essence of MI is the solution to an inverse problem using the data of stress patterns on the surface of tissue compressed by a pressure sensor array. Imaged tissue structures are presented in terms of their viscoelastic properties. Evaluation of tissue ‘hardness’ (shear elasticity modulus) provides a means for characterizing the tissue, differentiating normal and diseased conditions and detecting tumors and other lesions. In contrast to the other existing methods of medical imaging which use sophisticated hardware such as superconductive magnets, expensive X-ray equipment and complex ultrasonic phased arrays, MI hardware consists of inexpensive mechanical sensors and a positioning system connected to a PC. A key feature of MI is ‘knowledge-based imaging’. To produce a three-dimensional image, the computer uses both the measured parameters of an individual examined object and a general database on anatomy and pathology of the object. Two applications of MI are currently being developed: MI for mass screening and detection of breast cancer and MI for imaging the prostate and diagnosing prostate diseases. A prototype of the device for mechanical imaging of the prostate has been developed and is being tested clinically at the Robert Wood Johnson Medical School, New Jersey. The device is comprised of a transrectal probe with a position sensor and a pressure sensor array mounted on the articulated tip, an electronic unit and a PC. Results of extensive laboratory studies with rubber prostate models and initial data obtained in clinical trials strongly suggests that for certain applications the MI technology, as a new modality of imaging, has a diagnostic potential comparable to that of conventional diagnostic technologies. Mechanical imaging of the prostate appeared to be an efficient means of objectively evaluating and imaging the prostate and detecting prostate cancer. © 1998 Elsevier Science Ireland Ltd. All rights reserved. Keywords: Mechanical imaging; Medical diagnostics; Pressure sensor array
1. Introduction Mechanical imaging (MI) is a new technology of medical diagnostics in which internal
structures are visualized by sensing the pattern of mechanical stresses on the surface of an organ [1,2]. MI is capable of ‘seeing’ inside biological tissues without using any form
1386-5056/98/$19.00 © 1998 Elsevier Science Ireland Ltd. All rights reserved. PII S1386-5056(98)00040-9
196
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
of radiation. The most promising applications of MI devices are in those fields of medicine where palpation is proven to be a sensitive tool in detecting and monitoring diseases, namely breast and prostate cancer. Palpatory self-examination, widely advised and taught to women as a means of preclinical testing, contributes significantly to early cancer detection. In the review paper Control of breast cancer through mass screening by P. Strax [3], it is noted that ‘‘over 90% of breast cancer is first detected by women themselves, who bring the problem to their physicians.’’ The usefulness of palpatory self-examination as a preclinical test is fully proven by a wealth of data. Although not every breast cancer is accompanied by palpable structural changes in the breast tissue, a significant fraction of cancers do result in changes in tissue hardness. The fraction of these palpable cancers is large enough to justify efforts towards developing instrumental means to measure more objectively and quantitatively the same mechanical information as that obtained by a skillful physician using his fingers. Mechanical imaging has the potential for accomplishing this task. Another obvious area of application of MI technology is prostate cancer detection. Actually, most efforts on the development of MI were made on this application and the main focus of this paper will be on mechanical imaging of the prostate. Prostate carcinoma is the second leading cause of death from cancer among men; 25% of men with prostate cancer will die of the disease [4]. In a disease of such significance screening methods for early detection are pertinent. The armamentarium of today’s tools considered for such a task consists of prostatic specific antigen (PSA), digital rectal examination (DRE) and transrectal ultrasound (TRUS). PSA alone is the most sensitive single test for prostate cancer detection but its positive predictive
value is only 31% [5]. However 20–25% of patients with normal PSA who have prostate cancer will fall through the screening filter, although some of them may be detected by DRE. The DRE alone with its predictive value of 22% is even less useful, however, combining the two will boost the cancer detection rate nearly twice [5]. The combination has been recommended by the AUA and the American Cancer Society for screening and recently approved by the FDA for this purpose in the patients between the ages of 50 and 75 years. While the PSA levels provide an objective assessment with some overlaps in ranges between cancerous and non-cancerous glands, the DRE is dependent on the experience of the examiner. Despite the obvious usefulness of the diagnostic information obtained by DRE, there are no technical means or devices capable of yielding data similar to that obtained by the finger of a skilled examiner. The palpatory findings are in accordance between experienced urologists in only 84% of cases [6]. The disagreement among inexperienced examiners is higher and they will be most likely to carry the bulk of the screening. Once a lesion is palpated, the documentation depends on the precision of a physician’s description or diagram drawing. A state-of-the-art biomechanical measurement will obviate the flaws of human error with objective and quantitative imaging and documentation with high precision and resolution. It will also obviate the need for an experienced examiner in its final stage, as the examination is guided by computer. The data with diagnostic suggestions will then be presented to the physician for final approval or disapproval. A study done by [7] showed that DRE is the most cost efficient prostate screening method, especially if combined with prostatespecific antigen (PSA) methods. The study however points out that the effectiveness and
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
reliability of DRE is highly dependent on the level of skill of an examiner, since the finger as an instrument does not provide any quantitative information. The examiner instinctively relates what he senses by his finger to his previous experience with DRE and there may not be sufficient numbers of available skilled examiners required for large scale mass prostate screening. The mechanical imaging device may eliminate these limitations by providing an objective and quantitative method of measuring the elasticity of prostate tissues in combination with geometrical features of the prostate, while producing 3-D images using both the measured parameters of an individual prostate and a general database on prostate anatomy and pathology. There were many attempts to develop methods and devices for sensing regions of hardening in tissues and thus mimicking manual palpation for detection of cancer [8,9]. Although these multiple attempts were not very successful, the efforts of the researchers were justified by the fact that there is no physical property of soft biological tissues other than Young’s or shear moduli that can provide a great contrast between normal and diseased tissues [10]. The very limited success of these prior attempts to mimic palpation by devices detecting stress or strain patterns in the compressed breast can be explained by a number of reasons, the most important of which is the underestimation of the role of the examiner’s brain in the manual examination process. Palpation skills involve a complex analysis of temporal and spatial variation of force exerted on the tip of a finger, it is highly dependent on the data processing in the brain of the experienced examiner, on his ability to visualize sensory information. Processing the data in the MI method presented in this work simulates a skilled exam-
197
iner’s brain when detecting tumors by palpation analyzing the relative, rather than absolute, values of pressure and measuring temporal and spatial variation of the signals from the pressure sensor array. Intuitively, it may seem, that one should aim at acquiring precise data from the sensors under ideal conditions when all other mechanical signals, such as shaking of the operator’s hand or the beating heart and breathing of the patient, are eliminated. In reality, the opposite is true. A ‘noise’ that could modify the contribution of the lesion to the pressure pattern may serve as a means to detect and amplify this contribution. Consequently, one may deliberately introduce additional mechanical disturbance into the measuring process, such as oscillating, shifting, or tilting the probe, so the resulting spatial–temporal dependencies of the pressure pattern may serve as a source of information on the presence and parameters of the lesions. This principle is illustrated in Fig. 1 showing a pressure-sensing array pressed against the breast and then periodically moved over the rib cage. The measured pattern of pressure contains information on the mechanical heterogeneity of breast as well as on the geometry of the underlying rib cage. The contribution of the ribs into the spatial and temporal variations in the pressure pattern is much higher than that from soft tissue heterogeneity. The signal from the ribs may be considered as ‘noise’, but at the same time, the presence of the ribs provide an efficient mechanism for enhancing and detecting the signal from a lump in the breast. Within the breast, any lump that moves together with the surrounding tissue over the ribs produces an additional periodic stress on the individual sensors in the array. Fig. 1a and Fig. 1b show cross sections of a breast indicating the direction of movement transverse to the ribs and the parameters used for calculations of the dependencies presented in
198
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 1. Enhancements of the signal from a hard nodule in the breast by moving a pressure sensing array over the rib cage (see text).
Fig. 1c. Fig. 1c graphically shows the changes in pressure profile after a lateral shift of the probe [1]. The calculations are made using the classical solution of Goodier [11] of the problem of compressing a medium containing a spherical or cylindrical inclusion. The ordinate DP/P0 represents the normalized change in the pressure profile, where DP is the difference between pressure profiles for the breast tissue with and without a nodule (‘tumor’). The graph clearly shows how significant is the increase of the contribution of the tumor in the pressure profile after lateral shift. The ribs play the role of an amplifier of the measured effect. Temporal and spatial variations of the signals from the sensors contain information on mechanical heterogeneity of breast tissue. A key feature of our approach is ‘knowledge-based imaging’ [12], the processing of measured data based on biomechanical information that includes established correlation between mechanical, anatomical and histopathological properties of tissue as well as specific data for a particular patient. The ‘brain’ of our device will comprise a biomechanical model of the examined object and examination process. Yet another feature of our approach is adaptive data processing which enables the measuring system to ‘learn’ specific properties
and peculiarities of a particular investigated object and adjust the model to reflect unique characteristics of the investigated object [13,14].
2. Detection of hard nodules in soft tissues. Laboratory model studies Extensive theoretical studies on the inverse problem of mechanics (reconstruction of internal mechanical structure of an elastic object from the measurements of the surface strain and stress patterns), design and construction of various devices for measuring viscoelastic properties of soft tissues and biomechanical studies of surgically excised normal and diseased tissue samples preceded the development of the MI technology. Most of the laboratory studies aimed at proving the feasibility of the MI were conducted with the use of breast and prostate models. Models (tissue mimicking ‘phantoms’) were made of a polymer material consisting of a liquid plastic combined with either a softener or a hardener. By varying the proportion of these two components, it is possible to produce composite phantoms of any desired elasticity. Phantoms of breast and prostate of different shape, dimensions and hardness, with and without inclusions, were manufactured. Fig. 2
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
199
Fig. 2. Rubber models of the prostate gland representing the following conditions: normal prostate, BPH and various stages of cancer.
shows some of the developed life-like rubber models simulating normal prostate, BPH and various stages of prostate cancer. The phantoms simulate geometrical features of the specific state of the prostate, as well as the mechanical properties of the gland itself and palpable nodules. Fig. 3 presents some of the breast phantoms studied and an example of a rectangular rubber test-block with an inclusion. Initial theoretical estimates of the sensitivity of MI and its potential in detecting hard nodules in soft tissues made under certain assumptions on the boundary conditions yielded very promising results. Fig. 4 presents such an estimate of the sensitivity of MI for detecting a lesion in tissue in comparison with conventional palpation [1]. The figure illustrates the changes in the pressure profile ratio (DP/P0) at the point of the surface above the tumor as a function of the diameter d of the tumor at the depth h equal to 10
mm. Elasticity moduli ratios E/E0 (tumor/ normal) are taken equal to 5.0 (bold curve) and 2.0 (thin curve). Threshold level of sensitivity (dashed horizontal line) indicates that, ideally, the detection of a tumor 3 mm in diameter is possible, which is far superior to the manual palpation threshold. Palpation typically senses lumps of about 8–10 mm in diameter. Such an estimate for an ideal homogeneous system with oversimplified boundary conditions may give only an order of magnitude estimate of the sensitivity of the MI. This simulation does not take into account many real factors of manual palpation, such as the dynamic modes of examination and does not consider the scale of the natural heterogeneity of normal tissues surrounding the nodule. Palpation, similar to other sensory systems connected to the brain, senses the relative rather than absolute value of a variable to be evaluated. Sensing just a change of a measur-
200
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 3. Breast phantoms and an example of a rubber test-block with an inclusion.
able parameter, i.e. its spatial and temporal derivatives, is a basic principle of physiology of sensory perception. Fig. 5 presents results of an experiment on detecting hard inclusion in rubber models from the temporal and spatial features of a pressure pattern on the surface of the model. To be able to evaluate simultaneously both temporal and spatial profiles of pressure, a rectangular rubber phantom with an inclusion was placed over the pressure sensor array and pressure patterns were measured every 10 ms while a roller was moved over the model, as shown in Fig. 5A. Fig. 5B shows temporal pressure profiles for the sensors situated at different
Fig. 4. Theoretical estimation of the sensitivity of MI for detecting a lesion in tissue in comparison with conventional palpation.
distances from the inclusion (the distance for each profile is given at the right hand side of the figure). One can clearly see how the profiles differ depending on the relative position of the sensor and the nodule. A number of temporal and spatial features of the signal can be used to design algorithms for detecting the presence of a nodule: (a) the amplitude of the signal; (b) the width of the peak; (c) the symmetry of the peak; (d) the characteristic shape of the pressure profile, such as double hump profile, etc. Each of these algorithms has different potentials in extracting the useful part of the signal and discriminating various sources of noise, such as heterogeneity of tissue. Characteristic features of the pressure profiles shown in Fig. 5 result from a specific dynamic behavior of a nodule under the stress caused by a moving roller. When the roller approaches the nodule, the latter is pushed forward in the direction of motion of the roller. As soon as the roller reaches a position above the nodule, it ‘jumps’ in the opposite direction, producing characteristic distortions of the signals shown in the figure. The same phenomenon occurs, when a finger of the physician moves over a nodule in the breast during palpation. Such dynamic responses are the key features used by the brain of an experienced examiner to discriminate the signal from the nodule hidden in the background noise caused by tissue heterogeneity.
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
201
Fig. 5. Spatial-temporal pressure profiles on the surface of the phantom with an inclusion for the sensors situated at different distances from the inclusion (the distance for each profile is given at the right hand side of figure B).
One of the important parameters of the device should be its capability to determine the location of the nodule. Fig. 5C shows the data of Fig. 5B as the simplest topographic map where the levels of amplitude of the signal are displayed as a function of time and coordinate. One can see that even without any mathematical processing the center of gravity of the rings in the topographic image of Fig. 5C can be estimated with the accuracy of better than 1 mm in the horizontal plane. One of the problems in designing the pressure sensing array is the choice of the number and dimensions of individual sensors in the array. This choice is important for making a trade-off between sensitivity and spatial resolution of the device. We have investigated the problem of the spatial resolution of the system as a function of the size of the single sensor in the array. Fig. 6 shows results of an experiment similar to that described in the previous figures but performed with sensors
of different size. The figures below the 3-D graphs show the area of an individual sensor in the array. The smaller the sensor, the higher the level of noise, for two reasons: a smaller sensor provides weaker signal and it is more sensitive to the spatial variations of the pressure. Bigger sensors integrate the signal obtained from larger area and there is a trade-off between the quality of the signal and the spatial resolution. Importantly, the loss of spatial resolution is insignificant; the figure on the right shows that even a sensor with an area of 1 cm2 clearly ‘feels’ lateral displacement of 1 mm. Fig. 6 clearly shows that the spatial resolution in the dynamic modes of measurement can be over an order of magnitude higher than the characteristic dimension of a single sensor. Spatial resolution is closely related to the temporal resolution: in the range of variables discussed in Fig. 6 the more frames of pressure pattern per unit of displacement are made, the higher is the spatial resolution.
Fig. 6. Spatial-temporal pressure profiles for the sensors situated at different distances from the inclusion measured by the sensors of different size. The figures below 3-D graphs show the area of an individual sensor in the array.
202 A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
203
Fig. 7. An example of a dynamic examination mode with the use of the ‘learning’ stage during data acquisition.
Fig. 7 illustrates another important feature of the dynamic mode of measurement, a feature that is similar to the palpation data processing made in the brain of a skilled examiner. Fig. 7 presents an example of a dynamic mode of measurement with the use of the ‘learning’ stage during data acquisition. The pressure pattern over the surface of a rubber breast model with a spherical inclusion has been measured using a pressure sensor array. During the process of compression along the Z axis 500 frames of pressure pattern were measured. The ‘learning’ of the device was done using the initial 50 frames of pressure pattern during the compression. A possibility to use a fraction of the examination process as a ‘learning’ stage results from the differences in the behavior of the medium regions with and without inclusion under different levels of compression. In this case, the mechanism of discrimination between the signals from the inclusion versus surrounding material is related to differences in the mechanical nonlinearity of various regions of a heterogeneous specimen. Fig. 7b–e shows successive patterns of a calculated parameter characterizing nonlinearity of the successive changes of the local pressures over the surface of the breast phantom. The patterns correspond to different stages of the compression process following the ‘learning’
stage. The lower dark region corresponds to the nipple and upper dark spot represents the nodule. It is clearly seen that the rates of the increase of the ‘nonlinearity parameter’ over these two characteristic regions of the image are significantly different: the inclusion emerges abruptly on the image upon increasing compression. Consequently, this rate serves as an efficient discrimination parameter for detection of inclusions. The results of model experiments, some of which were described above and the results of theoretical analysis of optimum algorithms for detection of tumors in the soft tissue using the data from pressure sensing arrays [15,16] strongly suggest that the potential of MI in detecting hard nodules could be much superior to that of manual palpation.
3. Method and device for mechanical imaging of the prostate The key objective of the development of the prostate MI device is to increase significantly the efficacy of cancer diagnosis and decrease the number of unnecessary biopsies. The device comprises a transrectal probe connected to a computer. A position sensor and an array of pressure sensors are mounted on the tip of the probe. The pattern of mechani-
204
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 8. The flow-chart illustrating the general concept of the prostate MI device and its functioning.
cal stress and its changes as a function of applied pressure and time provide comprehensive information on the internal mechanical properties and geometry, which, in turn, are evaluated from the pattern by solving the inverse problem [17]. The flow chart shown in Fig. 8 schematically illustrates the general concept of the device and its functioning. After inserting the transrectal probe into the rectum and pressing the ‘start’ button, the positioning system begins recording the relative 3-D coordinates of the sensing tip of the probe, which is moved over the prostate by an operator. The computer provides feedback information on the stage and completeness of the examination process. Mechanical and geometrical properties of the prostate calculated from the measured data, in combination with all relevant data from the database and personal records of the patient, are used to generate both the computer-aided preliminary diagnosis and the 3-D image of the investigated organ. Fig. 9 illustrates the principle of operation of the MI probe, with the moveable tip comprising a position sensor and the pressure sensor array, which faces the prostate region.
Upper left panel shows a schematic of the inserted probe, while panels at the lower row show select frames obtained by moving a linear array of pressure sensors. Upper right panel represents a reconstructed prostate. Fig. 10A shows the relationship of the probe, rectal wall and a prostate with a nodule in cross section. Fig. 10B illustrates the virtual lines of equal pressure calculated from the data obtained by the position sensor and pressure sensor array. Equal pressure lines denoted in Fig. 10B by numerals 1, 2, 3 and 4 are related to the virtual strain profile. The contour of the prostate shown in Fig. 10B by
Fig. 9. The principle of operation of the MI probe.
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
205
Fig. 10. (A) Relative position of the probe, rectal wall and a prostate with a nodule shown in cross section. (B) illustrates the virtual lines of equal pressure calculated from the data obtained by the position sensor and pressure sensor array.
the bold dotted line 5 is reconstructed using the equal pressure profile data and the nonlinearity of the strain/stress relationship. The data shown schematically in Fig. 10B are used to evaluate the virtual patterns of both strain and stress. The device for mechanical imaging of the prostate consists of a transrectal probe connected to a computer via an interface electronic unit. The transrectal probe (Fig. 11) comprises a moveable tip which contains an array of pressure sensors and a position/orientation sensor (sensing element of the 3SPACE® INSIDETRAK™ position/orientation tracking device made by Polhemus, Colchester, VT). The resolution of the 3-D position measurements achievable with this particular system is 0.1 mm, assuming that the maximum distance between the transmitter (which is mounted in a belt and fixed at the back of the patient) and the sensing element mounted inside of the tip of the probe is no more than 50 cm. The pressure sensing array comprises six bending beam force sensors mounted inside the tip of the probe. The sensing part of each bending beam protrudes through the hole in the tip of the probe and
has force measuring area of 2 mm in diameter. When the tip of the probe touches the tissue, beams bend, which is detected by a pair of strain gauges bonded on opposite sides of each beam. Typical range of forces measured by each sensor is from 1 to 15 g with an accuracy of 0.2 g. The probe comprises also the ‘End of task’ button to communicate with the processor and a button for changing the position of the articulated tip of the probe. The tip of the probe is mated to a rigid tube, which in turn is attached to a pistol grip handle. A disposable rubber sheath covers the entire tip as well as the tube. Fig. 12 is a schematic diagram showing the general concept of the device and its function. The examination is conducted in accordance with a special protocol, describing a predetermined sequence of steps necessary to collect sufficient data. The computer provides feedback information on the stage and completeness of the examination process. Examination of the patient starts with inserting the probe into the rectum to the neighborhood of the prostate. At the moment when pressure sensors are passing the anus,
206
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 11. General view of the transrectal MI probe.
the processor automatically starts recording position and pressure data from the probe. The computer displays an image of typical pelvic anatomy and the position of the probe
Fig. 12. A schematic diagram showing general concept of the device and its function.
in relation to the anatomy along with a designation of the region to be scanned with the probe. This initial image serves as a map to navigate the probe. A cursor on the screen represents the tip of the examining probe. At this stage the relative positions of the probe and the prostate shown on the screen are a rough guess based on the ‘average’ male anatomy. This stage of examination is called ‘Positioning’. During this stage the operator moves the probe over the marked region to be scanned on the navigation map. The pressure exerted by the operator and the rate of movement of the probe is indicated on the screen to direct and optimize operator performance. The color of the cursor changes to indicate whether the pressure being applied is insufficient, excessive or adequate to optimally image the prostate. If the probe is being moved too fast, a warning signal is placed above the image on the screen. Similarly, if the probe is misoriented so that less than all the sensors contact tissue, an appropriate misorientation message is displayed. The aim of the ‘Positioning’ stage is to determine the exact location of the prostate in the test subject using accumulated position and pressure data and to produce an image of the prostate in the correct relation to the probe. At the end of the ‘Positioning’ stage the operator depresses a switch on the probe control handle sending an ‘End of task’ signal to the processor. Upon receiving the sig-
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
207
Fig. 13. Data collection phase of the prostate MI examination.
nal, the processor calculates the position of the center and boundaries of the prostate, thereby permitting the image on the screen to correctly show the relative position of the probe with respect to the actual prostate. These calculations take about 10–15 s, after which the image on the screen zooms out to reveal the prostate and neighboring structures. After this enlarged image of the prostate appears on the screen the process moves into the next phase. The processor program displays a sequential list of regions of the prostate to be examined adjacent to the image screen (Fig. 13) and the operator performs this sequence of operations. Advantageously, the choice of this sequence is similar to that of a regular digital rectal examination so that it is familiar to the operator. After completion of each
step of the sequence, the operator sends an ‘End of Task’ signal and the processor can check the task on the display and highlight the next task in the sequence. This data collection takes 2–3 min, after which the operator can terminate the examination by pressing the ‘End of task’ button and removing the probe from the patient. In the final step, the computer calculates a virtual pressure pattern, then based on this pattern, calculates a mechanical model of the prostate and finally, generates a three dimensional image of the examined prostate. It displays the examined prostate and in parallel, displays a reference ‘normal’ prostate. The operator is able to rotate the image of the reconstructed prostate on the screen, by using an associated computer mouse. The computer can synchronously rotate the refer-
208
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
ence prostate to facilitate comparison and detection of abnormalities. Fig. 14 presents a flow chart for data acquisition and processing in the device. The set of real time force data obtained from each sensor is processed into a set of transformed data. By elementary geometrical processing, the actual position data permits determination of three spatial coordinates for each pressure sensor. Thus, one can relate every measured force value from each sensor to a point in three-dimensional space. The force
Fig. 14. A chart showing the flow of data in the prostate imaging process.
measured by the sensor and its three coordinates are the transformed data. The transformed data can now be processed to provide pressure field data. The transformed data contain both useful information about parameters of the tissue being investigated and noise of various origins. The noise originates from force and position measurement error and from artifacts related to tissue movement (movement of the prostate, movement of the patient). Pressure field data is calculated by processing the transformed data to minimize noise and extract the 3-D spatial distribution of pressure approximating ideal conditions of measurement. While there are a number of possible algorithms for this processing, our initial approach is to use Chebyshev approximation [18]. The pressure field is represented as a superposition of Chebyshev polynomial functions. From the pressure field data approximated by Chebyshev polynomials, one can now reconstruct an elasticity model of the prostate —a model that will show the distribution of hardness in three-dimensional space. The surface of the examined prostate can be obtained by choosing a level of force corresponding to deformation of the rectal wall that permits the sensors to press against the prostate surface. From the pressure gradients, information on the prostate tissue hardness can be generated. One of the key tasks of the data processing in the MI method is to relate the data on the spatial distribution of the elasticity of the prostate to a database of general and patient specific medical data. In traditional medical imaging, the device usually displays the structure of an object in terms of some measured physical property. The image obtained this way is often very far from what the actual examined region of body or an organ would look like if exposed to direct sunlight, or
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
drawn by an artist. Therefore, an expert in a particular type of image analysis is required to tell the physician what information from the image is relevant to the diagnosis. Currently, as a result of a wider use of powerful computer means and databases, an alternative approach to imaging, so called knowledge-based imaging has started to emerge. Using knowledge-based imaging, a computer can store in memory a 3-D picture of a ‘normal’ prostate which is being examined and adjust (transform) this image according to the measured data, to produce an image that represents the actual examined gland. Such a pictorial 3-D image or its cross sections will additionally include data on the mechanical properties of the prostate. It will be significantly easier for a physician to recognize abnormalities of the examined organ, represented on such an image. Further, the expert system will use the knowledge about characteristics of different types and stages of prostate cancer to point out any poorly defined and suspicious regions in the model, or just show any abnormalities or deviations from what the ‘normal’ prostate should look like. At this point, the physician can also enter into the computer new information based on other tests or exams performed on the same prostate and the knowledge base will ‘learn’ and ‘expand’. Once a 3-D model of the actual examined prostate is stored in the computer, it must be presented by the user in a way that would allow both external and internal features to be seen on one picture. This means that the 3-D image on the screen should contain information about geometrical features of the prostate as well as spatial distribution of elasticity and surface texture information. Additionally, the image should indicate to the user which areas of the examined prostate are poorly defined and need to be examined further in order to produce a complete diag-
209
nosis. There are several potential 3-D visualization methods that can be suitable for this task, such as polygon based surface methods, ray cast volume rendering and cross section slicing. In order to utilize knowledge-based imaging methods in our project, we need to have a 3-D model of prostate in the computer which can be easily redimensioned to fit our measured data. For the internal computer representation of such modeled surfaces of the prostate, parametric spline surfaces (X(s, t), Y(s, t), Z(s, t)) comprised of triangular surface patches which are spliced together so that we have an overall smooth surface can be used. Each of these surfaces has a fixed topology in that the number and connectivity of individual triangular patches does not change but each has control point parameters which are used to control the shape and geometry of the surface. The actual values of the control point parameters are eventually selected so as to ‘best fit’ the data of a particular gland, but the number of surfaces and the number of triangular patches comprising each surface remains fixed (unless the user intervenes and requests a change).
3.1. Software for the prostate MI de6ice The software of the device is implemented as a 32-bit application for Microsoft ® Windows NT™ operation system. The user interface module is developed as a main EXE module and all calculation modules are implemented as a Windows function in a DLL. For the user interface module development we used the visual programming environment for rapid application development (RAD) Delphi 2.0™ and the program language Object Pascal. The user interface can provide two windows: the main window consisting of control elements which allow the user to
210
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
manage the process of examination and the visualization window for 3-D or 2-D representation of the imaged prostate. Key functions of the software are numerical filtration of the pressure and coordinate data flow and reconstruction of the 3-D virtual pressure pattern characterizing both geometrical and mechanical structure of the examined prostate. Reconstruction of the 3D virtual pressure pattern is made with the use of Chebyshev polynomials. The approximated function is given as a linear superposition of the basis functions where products of Chebyshev polynomials from spatial coordinates are used. The approximation parameters are the coefficients in such a linear combination. The coefficients are determined as a solution of a linear system of equations. The computer has in memory a 3-D model of a ‘normal’ prostate and adjusts (transforms) this model according to the measured data to produce an image that represents the actual examined organ of a particular patient. A number of potential visualization methods capable of displaying a realistic 3-D image of a prostate on the screen were evaluated. The decisions on the adequacy of a method for our objectives were based on its ability to not only display the geometrical features of the prostate, but also to indicate the spatial hardness and texture distribution in a clear and easily recognizable way. The tasks on pictorial representation of structural and mechanical features of the prostate were accomplished and a 3-D computer model of normal and diseased prostates were developed in collaboration with Dr. Marsha Jessup, Director of the Media Resources Department, Robert Wood Johnson Medical School, New Jersey. Fig. 15 shows a 3-D model of a normal prostate and a series of 3-D models of a prostate with two benign tumors near the urethra and a growing car-
Fig. 15. Three-dimensional computer model of a normal prostate and a series of 3-D models of a prostate with two benign tumors near the urethra and a growing carcinoma in the outer area of the prostate.
cinoma in the outer area of the prostate. In addition, 3-D computer models of the male urogenital system and rectum anatomy have been developed (Fig. 16) as a part of the anatomical database which will be needed in the specialized software and interactive diagnostic algorithms in the knowledge-based mechanical imaging device.
Fig. 16. Three-dimensional computer model of the male urogenital system and rectum anatomy.
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
The 3-D prostate reconstruction from the experimental data is based on the Finite Elements Method (FEM) widespread in CAD/ CAM systems to provide maximum reality in the displayed model. The model includes information on geometry as well as information on the tissue elasticity and possible nodules and is based on the use of isoparametric 3-D elements having 21 nodes and shelled 3-D isoparametric elements using eight nodes. A special fast algorithm of photorealistic image reconstruction has been developed. Brightness of different sections of the image surface depends on their inclination and takes into consideration both scattered and mirror-image components of reflection. The resources of the True Color palette are fully used. Using knowledge-based imaging methods requires having a 3-D model of a prostate in the computer that can be easily redimensioned to fit measured data. For the internal computer representation of such modeled surfaces of the prostate, parametric spline surfaces (X(s, t), Y(s, t), Z(s, t)) comprised of triangular surface patches which are spliced together to create an overall smooth surface can be used. Each of these surfaces has a fixed topology in that the number and connectivity of individual triangular patches does not change, but each has control point parameters which are used to control the shape and geometry of the surface [19]. The actual values of the control point parameters are eventually selected so as to ‘best fit’ the data of a particular gland, but the number of surfaces and the number of triangular patches comprising each surface remains fixed. To analyze quantitatively the spatial resolution of the device, to express accuracy and tolerance in millimeters and to be able to compare performance of the MI method with
211
that of DRE, a pilot experimental study has been performed. Fig. 17 shows cross-sections of the measured (upper row) and reconstructed (lower row) model prostates. Solid lines in the upper row figures represent actual profiles of the examined models and experimental points shown in the same figures represent filtered data. Fig. 17A shows the data obtained by sensing a prostate phantom by the MI probe directly and the data of Fig. 17B represents a similar experiment on the same model with one difference: the phantom was covered by a 3 mm thick rubber layer simulating the rectal wall. These experimental studies demonstrated the following capabilities of the developed software: the influence of the ‘rectal wall’ on the evaluation of the prostate parameters is not significant; the data acquisition part of the device is capable of reproducing geometrical features of the prostate with tolerance in the 1–2 mm range; knowledge-based reconstruction algorithms do not add significant distortions in the simulated prostate 3-D model. Special experiments were performed to obtain a quantitative comparison of the MI and DRE evaluation of the geometrical features of the prostate models. A number of researchers and doctors were asked to palpate different prostate models through an intermediate layer of soft rubber and record their observations. Specific parameters including width and length of the model, depth of median groove and asymmetry of the model were noted. The MI device was then used to examine the models. The MI device was found to evaluate all geometrical parameters about five times more accurately than the human finger. This means that in a simplified model situation were only geometrical features are considered, we can potentially have
212
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 17. Cross-sections of the measured and reconstructedmodel prostate: (A) without intermediate layer; and (B) covered by a 3 mm thick rubber layer.
five times improvement in examination accuracy when MI is used compared to DRE. Experiments on the prostate phantoms were performed to test developed algorithms of knowledge-based 3-D reconstruction of the prostate (Figs. 18 and 19). Fig. 18 shows experimental data on the ‘equal pressure’ patterns (see explanations for Fig. 10) obtained for a prostate phantom shown schematically on the right panel of the figure. The prostate phantom during the measurements was covered by 3 mm thick layer of soft rubber simulating the rectal wall. Fig. 19A shows calculated surface of the model using the ‘equal pressure’ patterns based on the data received from the force and position sensors. The 3-D pattern obtained enables quantitative evaluation of the geometrical features of the examined model such as dimensions, symmetry, presence and depth of median groove, etc. Fig. 19B is a result of knowledge based
3-D reconstruction of the prostate using the data of Fig. 19A. These experiments have shown that all the important and diagnostically relevant features of the prostate model are reconstructed without significant distortion.
4. Conclusions Results of the laboratory pilot studies have proven the feasibility of the mechanical imaging technology. The project on development of the first MI device for diagnosing prostate diseases and detecting prostate cancer has now moved into the phase of clinical trials. The Robert Wood Johnson Medical Center Institutional Review Board has approved the clinical testing of the prototype and a number patients were examined and 3-D mechanical images of in vivo prostates have been ob-
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
213
Fig. 18. Experimental data on the ‘equal pressure patterns’ for a prostate phantom shown schematically on the right panel of the figure.
tained. Clinical studies were conducted under the guidance of Professor K.B. Cummings and Dr R.E. Weiss. Here we present just two examples of the clinical examination data. A complete report on the results of the clinical studies will be presented elsewhere. Fig. 20 shows comparative projections of three-dimensional images of the first mechanically imaged prostate of a
healthy volunteer (right column) and a computer simulated ‘ideal’ prostate (left column). There were no abnormalities seen and the prostate appeared to have a uniform consistency. Fig. 21 presents the first MI data on a patient with histologically confirmed prostate cancer. This patient was noted to have a PSA of 11.1 (normal 1–4.0) and had a small nod-
214
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 19. Knowledge based three-dimensional reconstruction of the prostate phantom.
ule noted on digital rectal examination. Because of these findings, rectal sonography and random prostate needle biopsies were performed. The rectal sonography revealed an enlarged 58.8 g. prostate, which had no
ultrasonically visible abnormalities (upper right panel in Fig. 21). Random biopsies revealed a 2/5 Gleason score prostate cancer on the left side of the prostate, which correlates with the nodule palpated on DRE. MI data shown in Fig. 21 as 3-D virtual strain pattern (upper left panel) and a cross-section of the force/position data across the prostate apex (lower panel), clearly show the location and dimensions of the prostate cancer nodule missed by ultrasound. These promising results portend the creation of a simple and inexpensive device for imaging the prostate, evaluating its state and diagnosing prostate diseases, which utilizes physical principles and measured parameters similar to those associated with a digital rectal examination. Preliminary data obtained in the pre-clinical testing of the prostate MI device strongly suggest that mechanical imaging technology meets basic requirements for a mass cancer screening method and for an inexpensive and affordable method of day-to-day monitoring of cancer in its advanced stages: it is simple, fast, inexpensive and safe.
Fig. 20. Comparative projections of three-dimensional images of the first mechanically imaged prostate of a healthy volunteer (right column) and a computer simulated ‘ideal’ prostate (left column).
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
215
Fig. 21. Comparison of the MI and sonography data obtained for a patient with a palpable nodule in the left part of the prostate apex where the prostate cancer was revealed by random biopsy.
However, the MI method is general and has much broader implications in a variety of diagnostic and surgical applications that require cancer diagnosis and assessment and localization of abnormal tissue for biopsy and characterization. Acknowledgements This work was supported in part by the National Cancer Institute (USA) grant 1 R43 CA69175-01. References [1] A.P. Sarvazyan, A.R. Skovoroda, Method and apparatus for elasticity imaging, US Patent 5524636, 1996.
[2] A.P. Sarvazyan, Knowledge-based mechanical imaging. Proc. 10th IEEE Symp. on Computer Based Medical Systems’97, IEEE Computer Society publication, Washington, 1997, pp. 120 – 125. [3] P. Strax, Control of breast cancer through mass screening, Hospimedica March/April (1989) 35–40. [4] C.C. Bohring, T.S. Squires, T. Tong, Cancer statistic, CA Cancer J. Clin. 43 (1993) 7 – 26. [5] W.J. Catalona, J.P. Richie, F.R. Ahmann, et al., A multicenter evaluation of PSA and digital rectal examination (DRE) for early detection of prostate cancer in 6374 volunteers, J. Urol. 149 (Suppl) (1993) 412A. [6] D.S. Smith, W.J. Catalona, Inter-examiner variability of digital rectal examination in detecting prostate cancer, Urology 45 (1995) 70 – 74. [7] P.J. Littrup, A.C. Goodman, C.J. Mettlin, The benefit and cost of prostate cancer early detection, Cancer J. Clin. 43 (1993) 134 – 149. [8] E.H. Frei, B.D. Sollish, S. Yerushalmi, Instrument for viscoelastic measurement, U.S. Patent c 4,250,894, 1981.
216
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
[9] C.R. Gentle, Mammobarography: a possible method of mass breast screening, J. Biomed. Eng. 10 (1988) 124–126. [10] A.P. Sarvazyan, A.R. Skovoroda, S.Y. Emelianov, J.B. Fowlkes, J.G. Pipe, R.S. Adler, R.B. Buxton, P.L. Carson, Biophysical bases of elasticity imaging, in: J.P. Jones (Ed.), Acoust. Imaging, vol. 21, Plenum, New York, 1995, pp. 223 – 240. [11] J.N. Goodier, Concentration of stress around spherical and cylindrical inclusions and flaws, Trans. ASME 55 (39) (1933) 39–44. [12] A.P. Sarvazyan, F.L. Lizzi, P.N.T. Wells, A new philosophy of medical imaging, Med. Hypotheses 36 (1991) 327–335. [13] Y.P. Pyt’ev, Measurement computer systems of superhigh resolution, Pattern Recognit. Image Anal. 1 (1) (1991) 54–76. [14] Y.P. Pyt’ev, Fuzzy uncertain computer-aided measuring systems, Pattern Recognit. Image Anal. 4 (1994) 178–184.
.
[15] D.A. Pashko, Y.P. Pyt’ev, A.P. Sarvazyan, A.R. Skovoroda, On the ultimate potentialities of classification of pathologies in the problem of diagnostics of breast cancer, Pattern Recognit. Image Anal. 6 (1996a) 510 – 522. [16] D.A. Pashko, Y.P. Pyt’ev, A.P. Sarvazyan, Minimax evaluation of parameters of a nodule in diagnosing breast cancer with the use of a force sensor array, Bulletin of Moscow State University, Ser. 3, Physics, Astronomy (1996b) 18 – 25. [17] A.R. Skovoroda, Inverse problems of the theory of elasticity in the diagnostics of soft tissue pathologies, Preprint, Pushchino Research Center of the Russian Academy of Sciences, Pushchino (1992). [18] J.P. Boyd, Chebyshev and Fourier Spectral Methods, Springer, New York, 1989. [19] W. Welch, A. Witkin, Free-form shape design using triangulated surfaces, Comput. Graph. 28 (1994) 247 – 256.
Mechanical imaging: A new technology for medical diagnostics Armen Sarvazyan Artann Laboratories, 1 Ri6a A6enue, North Brunswick, NJ 08530 -3302, USA
Abstract Mechanical imaging (MI) is a newly developed modality of medical diagnostics based on reconstruction of tissue structure and viscoelastic properties using mechanical sensors. The essence of MI is the solution to an inverse problem using the data of stress patterns on the surface of tissue compressed by a pressure sensor array. Imaged tissue structures are presented in terms of their viscoelastic properties. Evaluation of tissue ‘hardness’ (shear elasticity modulus) provides a means for characterizing the tissue, differentiating normal and diseased conditions and detecting tumors and other lesions. In contrast to the other existing methods of medical imaging which use sophisticated hardware such as superconductive magnets, expensive X-ray equipment and complex ultrasonic phased arrays, MI hardware consists of inexpensive mechanical sensors and a positioning system connected to a PC. A key feature of MI is ‘knowledge-based imaging’. To produce a three-dimensional image, the computer uses both the measured parameters of an individual examined object and a general database on anatomy and pathology of the object. Two applications of MI are currently being developed: MI for mass screening and detection of breast cancer and MI for imaging the prostate and diagnosing prostate diseases. A prototype of the device for mechanical imaging of the prostate has been developed and is being tested clinically at the Robert Wood Johnson Medical School, New Jersey. The device is comprised of a transrectal probe with a position sensor and a pressure sensor array mounted on the articulated tip, an electronic unit and a PC. Results of extensive laboratory studies with rubber prostate models and initial data obtained in clinical trials strongly suggests that for certain applications the MI technology, as a new modality of imaging, has a diagnostic potential comparable to that of conventional diagnostic technologies. Mechanical imaging of the prostate appeared to be an efficient means of objectively evaluating and imaging the prostate and detecting prostate cancer. © 1998 Elsevier Science Ireland Ltd. All rights reserved. Keywords: Mechanical imaging; Medical diagnostics; Pressure sensor array
1. Introduction Mechanical imaging (MI) is a new technology of medical diagnostics in which internal
structures are visualized by sensing the pattern of mechanical stresses on the surface of an organ [1,2]. MI is capable of ‘seeing’ inside biological tissues without using any form
1386-5056/98/$19.00 © 1998 Elsevier Science Ireland Ltd. All rights reserved. PII S1386-5056(98)00040-9
196
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
of radiation. The most promising applications of MI devices are in those fields of medicine where palpation is proven to be a sensitive tool in detecting and monitoring diseases, namely breast and prostate cancer. Palpatory self-examination, widely advised and taught to women as a means of preclinical testing, contributes significantly to early cancer detection. In the review paper Control of breast cancer through mass screening by P. Strax [3], it is noted that ‘‘over 90% of breast cancer is first detected by women themselves, who bring the problem to their physicians.’’ The usefulness of palpatory self-examination as a preclinical test is fully proven by a wealth of data. Although not every breast cancer is accompanied by palpable structural changes in the breast tissue, a significant fraction of cancers do result in changes in tissue hardness. The fraction of these palpable cancers is large enough to justify efforts towards developing instrumental means to measure more objectively and quantitatively the same mechanical information as that obtained by a skillful physician using his fingers. Mechanical imaging has the potential for accomplishing this task. Another obvious area of application of MI technology is prostate cancer detection. Actually, most efforts on the development of MI were made on this application and the main focus of this paper will be on mechanical imaging of the prostate. Prostate carcinoma is the second leading cause of death from cancer among men; 25% of men with prostate cancer will die of the disease [4]. In a disease of such significance screening methods for early detection are pertinent. The armamentarium of today’s tools considered for such a task consists of prostatic specific antigen (PSA), digital rectal examination (DRE) and transrectal ultrasound (TRUS). PSA alone is the most sensitive single test for prostate cancer detection but its positive predictive
value is only 31% [5]. However 20–25% of patients with normal PSA who have prostate cancer will fall through the screening filter, although some of them may be detected by DRE. The DRE alone with its predictive value of 22% is even less useful, however, combining the two will boost the cancer detection rate nearly twice [5]. The combination has been recommended by the AUA and the American Cancer Society for screening and recently approved by the FDA for this purpose in the patients between the ages of 50 and 75 years. While the PSA levels provide an objective assessment with some overlaps in ranges between cancerous and non-cancerous glands, the DRE is dependent on the experience of the examiner. Despite the obvious usefulness of the diagnostic information obtained by DRE, there are no technical means or devices capable of yielding data similar to that obtained by the finger of a skilled examiner. The palpatory findings are in accordance between experienced urologists in only 84% of cases [6]. The disagreement among inexperienced examiners is higher and they will be most likely to carry the bulk of the screening. Once a lesion is palpated, the documentation depends on the precision of a physician’s description or diagram drawing. A state-of-the-art biomechanical measurement will obviate the flaws of human error with objective and quantitative imaging and documentation with high precision and resolution. It will also obviate the need for an experienced examiner in its final stage, as the examination is guided by computer. The data with diagnostic suggestions will then be presented to the physician for final approval or disapproval. A study done by [7] showed that DRE is the most cost efficient prostate screening method, especially if combined with prostatespecific antigen (PSA) methods. The study however points out that the effectiveness and
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
reliability of DRE is highly dependent on the level of skill of an examiner, since the finger as an instrument does not provide any quantitative information. The examiner instinctively relates what he senses by his finger to his previous experience with DRE and there may not be sufficient numbers of available skilled examiners required for large scale mass prostate screening. The mechanical imaging device may eliminate these limitations by providing an objective and quantitative method of measuring the elasticity of prostate tissues in combination with geometrical features of the prostate, while producing 3-D images using both the measured parameters of an individual prostate and a general database on prostate anatomy and pathology. There were many attempts to develop methods and devices for sensing regions of hardening in tissues and thus mimicking manual palpation for detection of cancer [8,9]. Although these multiple attempts were not very successful, the efforts of the researchers were justified by the fact that there is no physical property of soft biological tissues other than Young’s or shear moduli that can provide a great contrast between normal and diseased tissues [10]. The very limited success of these prior attempts to mimic palpation by devices detecting stress or strain patterns in the compressed breast can be explained by a number of reasons, the most important of which is the underestimation of the role of the examiner’s brain in the manual examination process. Palpation skills involve a complex analysis of temporal and spatial variation of force exerted on the tip of a finger, it is highly dependent on the data processing in the brain of the experienced examiner, on his ability to visualize sensory information. Processing the data in the MI method presented in this work simulates a skilled exam-
197
iner’s brain when detecting tumors by palpation analyzing the relative, rather than absolute, values of pressure and measuring temporal and spatial variation of the signals from the pressure sensor array. Intuitively, it may seem, that one should aim at acquiring precise data from the sensors under ideal conditions when all other mechanical signals, such as shaking of the operator’s hand or the beating heart and breathing of the patient, are eliminated. In reality, the opposite is true. A ‘noise’ that could modify the contribution of the lesion to the pressure pattern may serve as a means to detect and amplify this contribution. Consequently, one may deliberately introduce additional mechanical disturbance into the measuring process, such as oscillating, shifting, or tilting the probe, so the resulting spatial–temporal dependencies of the pressure pattern may serve as a source of information on the presence and parameters of the lesions. This principle is illustrated in Fig. 1 showing a pressure-sensing array pressed against the breast and then periodically moved over the rib cage. The measured pattern of pressure contains information on the mechanical heterogeneity of breast as well as on the geometry of the underlying rib cage. The contribution of the ribs into the spatial and temporal variations in the pressure pattern is much higher than that from soft tissue heterogeneity. The signal from the ribs may be considered as ‘noise’, but at the same time, the presence of the ribs provide an efficient mechanism for enhancing and detecting the signal from a lump in the breast. Within the breast, any lump that moves together with the surrounding tissue over the ribs produces an additional periodic stress on the individual sensors in the array. Fig. 1a and Fig. 1b show cross sections of a breast indicating the direction of movement transverse to the ribs and the parameters used for calculations of the dependencies presented in
198
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 1. Enhancements of the signal from a hard nodule in the breast by moving a pressure sensing array over the rib cage (see text).
Fig. 1c. Fig. 1c graphically shows the changes in pressure profile after a lateral shift of the probe [1]. The calculations are made using the classical solution of Goodier [11] of the problem of compressing a medium containing a spherical or cylindrical inclusion. The ordinate DP/P0 represents the normalized change in the pressure profile, where DP is the difference between pressure profiles for the breast tissue with and without a nodule (‘tumor’). The graph clearly shows how significant is the increase of the contribution of the tumor in the pressure profile after lateral shift. The ribs play the role of an amplifier of the measured effect. Temporal and spatial variations of the signals from the sensors contain information on mechanical heterogeneity of breast tissue. A key feature of our approach is ‘knowledge-based imaging’ [12], the processing of measured data based on biomechanical information that includes established correlation between mechanical, anatomical and histopathological properties of tissue as well as specific data for a particular patient. The ‘brain’ of our device will comprise a biomechanical model of the examined object and examination process. Yet another feature of our approach is adaptive data processing which enables the measuring system to ‘learn’ specific properties
and peculiarities of a particular investigated object and adjust the model to reflect unique characteristics of the investigated object [13,14].
2. Detection of hard nodules in soft tissues. Laboratory model studies Extensive theoretical studies on the inverse problem of mechanics (reconstruction of internal mechanical structure of an elastic object from the measurements of the surface strain and stress patterns), design and construction of various devices for measuring viscoelastic properties of soft tissues and biomechanical studies of surgically excised normal and diseased tissue samples preceded the development of the MI technology. Most of the laboratory studies aimed at proving the feasibility of the MI were conducted with the use of breast and prostate models. Models (tissue mimicking ‘phantoms’) were made of a polymer material consisting of a liquid plastic combined with either a softener or a hardener. By varying the proportion of these two components, it is possible to produce composite phantoms of any desired elasticity. Phantoms of breast and prostate of different shape, dimensions and hardness, with and without inclusions, were manufactured. Fig. 2
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
199
Fig. 2. Rubber models of the prostate gland representing the following conditions: normal prostate, BPH and various stages of cancer.
shows some of the developed life-like rubber models simulating normal prostate, BPH and various stages of prostate cancer. The phantoms simulate geometrical features of the specific state of the prostate, as well as the mechanical properties of the gland itself and palpable nodules. Fig. 3 presents some of the breast phantoms studied and an example of a rectangular rubber test-block with an inclusion. Initial theoretical estimates of the sensitivity of MI and its potential in detecting hard nodules in soft tissues made under certain assumptions on the boundary conditions yielded very promising results. Fig. 4 presents such an estimate of the sensitivity of MI for detecting a lesion in tissue in comparison with conventional palpation [1]. The figure illustrates the changes in the pressure profile ratio (DP/P0) at the point of the surface above the tumor as a function of the diameter d of the tumor at the depth h equal to 10
mm. Elasticity moduli ratios E/E0 (tumor/ normal) are taken equal to 5.0 (bold curve) and 2.0 (thin curve). Threshold level of sensitivity (dashed horizontal line) indicates that, ideally, the detection of a tumor 3 mm in diameter is possible, which is far superior to the manual palpation threshold. Palpation typically senses lumps of about 8–10 mm in diameter. Such an estimate for an ideal homogeneous system with oversimplified boundary conditions may give only an order of magnitude estimate of the sensitivity of the MI. This simulation does not take into account many real factors of manual palpation, such as the dynamic modes of examination and does not consider the scale of the natural heterogeneity of normal tissues surrounding the nodule. Palpation, similar to other sensory systems connected to the brain, senses the relative rather than absolute value of a variable to be evaluated. Sensing just a change of a measur-
200
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 3. Breast phantoms and an example of a rubber test-block with an inclusion.
able parameter, i.e. its spatial and temporal derivatives, is a basic principle of physiology of sensory perception. Fig. 5 presents results of an experiment on detecting hard inclusion in rubber models from the temporal and spatial features of a pressure pattern on the surface of the model. To be able to evaluate simultaneously both temporal and spatial profiles of pressure, a rectangular rubber phantom with an inclusion was placed over the pressure sensor array and pressure patterns were measured every 10 ms while a roller was moved over the model, as shown in Fig. 5A. Fig. 5B shows temporal pressure profiles for the sensors situated at different
Fig. 4. Theoretical estimation of the sensitivity of MI for detecting a lesion in tissue in comparison with conventional palpation.
distances from the inclusion (the distance for each profile is given at the right hand side of the figure). One can clearly see how the profiles differ depending on the relative position of the sensor and the nodule. A number of temporal and spatial features of the signal can be used to design algorithms for detecting the presence of a nodule: (a) the amplitude of the signal; (b) the width of the peak; (c) the symmetry of the peak; (d) the characteristic shape of the pressure profile, such as double hump profile, etc. Each of these algorithms has different potentials in extracting the useful part of the signal and discriminating various sources of noise, such as heterogeneity of tissue. Characteristic features of the pressure profiles shown in Fig. 5 result from a specific dynamic behavior of a nodule under the stress caused by a moving roller. When the roller approaches the nodule, the latter is pushed forward in the direction of motion of the roller. As soon as the roller reaches a position above the nodule, it ‘jumps’ in the opposite direction, producing characteristic distortions of the signals shown in the figure. The same phenomenon occurs, when a finger of the physician moves over a nodule in the breast during palpation. Such dynamic responses are the key features used by the brain of an experienced examiner to discriminate the signal from the nodule hidden in the background noise caused by tissue heterogeneity.
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
201
Fig. 5. Spatial-temporal pressure profiles on the surface of the phantom with an inclusion for the sensors situated at different distances from the inclusion (the distance for each profile is given at the right hand side of figure B).
One of the important parameters of the device should be its capability to determine the location of the nodule. Fig. 5C shows the data of Fig. 5B as the simplest topographic map where the levels of amplitude of the signal are displayed as a function of time and coordinate. One can see that even without any mathematical processing the center of gravity of the rings in the topographic image of Fig. 5C can be estimated with the accuracy of better than 1 mm in the horizontal plane. One of the problems in designing the pressure sensing array is the choice of the number and dimensions of individual sensors in the array. This choice is important for making a trade-off between sensitivity and spatial resolution of the device. We have investigated the problem of the spatial resolution of the system as a function of the size of the single sensor in the array. Fig. 6 shows results of an experiment similar to that described in the previous figures but performed with sensors
of different size. The figures below the 3-D graphs show the area of an individual sensor in the array. The smaller the sensor, the higher the level of noise, for two reasons: a smaller sensor provides weaker signal and it is more sensitive to the spatial variations of the pressure. Bigger sensors integrate the signal obtained from larger area and there is a trade-off between the quality of the signal and the spatial resolution. Importantly, the loss of spatial resolution is insignificant; the figure on the right shows that even a sensor with an area of 1 cm2 clearly ‘feels’ lateral displacement of 1 mm. Fig. 6 clearly shows that the spatial resolution in the dynamic modes of measurement can be over an order of magnitude higher than the characteristic dimension of a single sensor. Spatial resolution is closely related to the temporal resolution: in the range of variables discussed in Fig. 6 the more frames of pressure pattern per unit of displacement are made, the higher is the spatial resolution.
Fig. 6. Spatial-temporal pressure profiles for the sensors situated at different distances from the inclusion measured by the sensors of different size. The figures below 3-D graphs show the area of an individual sensor in the array.
202 A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
203
Fig. 7. An example of a dynamic examination mode with the use of the ‘learning’ stage during data acquisition.
Fig. 7 illustrates another important feature of the dynamic mode of measurement, a feature that is similar to the palpation data processing made in the brain of a skilled examiner. Fig. 7 presents an example of a dynamic mode of measurement with the use of the ‘learning’ stage during data acquisition. The pressure pattern over the surface of a rubber breast model with a spherical inclusion has been measured using a pressure sensor array. During the process of compression along the Z axis 500 frames of pressure pattern were measured. The ‘learning’ of the device was done using the initial 50 frames of pressure pattern during the compression. A possibility to use a fraction of the examination process as a ‘learning’ stage results from the differences in the behavior of the medium regions with and without inclusion under different levels of compression. In this case, the mechanism of discrimination between the signals from the inclusion versus surrounding material is related to differences in the mechanical nonlinearity of various regions of a heterogeneous specimen. Fig. 7b–e shows successive patterns of a calculated parameter characterizing nonlinearity of the successive changes of the local pressures over the surface of the breast phantom. The patterns correspond to different stages of the compression process following the ‘learning’
stage. The lower dark region corresponds to the nipple and upper dark spot represents the nodule. It is clearly seen that the rates of the increase of the ‘nonlinearity parameter’ over these two characteristic regions of the image are significantly different: the inclusion emerges abruptly on the image upon increasing compression. Consequently, this rate serves as an efficient discrimination parameter for detection of inclusions. The results of model experiments, some of which were described above and the results of theoretical analysis of optimum algorithms for detection of tumors in the soft tissue using the data from pressure sensing arrays [15,16] strongly suggest that the potential of MI in detecting hard nodules could be much superior to that of manual palpation.
3. Method and device for mechanical imaging of the prostate The key objective of the development of the prostate MI device is to increase significantly the efficacy of cancer diagnosis and decrease the number of unnecessary biopsies. The device comprises a transrectal probe connected to a computer. A position sensor and an array of pressure sensors are mounted on the tip of the probe. The pattern of mechani-
204
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 8. The flow-chart illustrating the general concept of the prostate MI device and its functioning.
cal stress and its changes as a function of applied pressure and time provide comprehensive information on the internal mechanical properties and geometry, which, in turn, are evaluated from the pattern by solving the inverse problem [17]. The flow chart shown in Fig. 8 schematically illustrates the general concept of the device and its functioning. After inserting the transrectal probe into the rectum and pressing the ‘start’ button, the positioning system begins recording the relative 3-D coordinates of the sensing tip of the probe, which is moved over the prostate by an operator. The computer provides feedback information on the stage and completeness of the examination process. Mechanical and geometrical properties of the prostate calculated from the measured data, in combination with all relevant data from the database and personal records of the patient, are used to generate both the computer-aided preliminary diagnosis and the 3-D image of the investigated organ. Fig. 9 illustrates the principle of operation of the MI probe, with the moveable tip comprising a position sensor and the pressure sensor array, which faces the prostate region.
Upper left panel shows a schematic of the inserted probe, while panels at the lower row show select frames obtained by moving a linear array of pressure sensors. Upper right panel represents a reconstructed prostate. Fig. 10A shows the relationship of the probe, rectal wall and a prostate with a nodule in cross section. Fig. 10B illustrates the virtual lines of equal pressure calculated from the data obtained by the position sensor and pressure sensor array. Equal pressure lines denoted in Fig. 10B by numerals 1, 2, 3 and 4 are related to the virtual strain profile. The contour of the prostate shown in Fig. 10B by
Fig. 9. The principle of operation of the MI probe.
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
205
Fig. 10. (A) Relative position of the probe, rectal wall and a prostate with a nodule shown in cross section. (B) illustrates the virtual lines of equal pressure calculated from the data obtained by the position sensor and pressure sensor array.
the bold dotted line 5 is reconstructed using the equal pressure profile data and the nonlinearity of the strain/stress relationship. The data shown schematically in Fig. 10B are used to evaluate the virtual patterns of both strain and stress. The device for mechanical imaging of the prostate consists of a transrectal probe connected to a computer via an interface electronic unit. The transrectal probe (Fig. 11) comprises a moveable tip which contains an array of pressure sensors and a position/orientation sensor (sensing element of the 3SPACE® INSIDETRAK™ position/orientation tracking device made by Polhemus, Colchester, VT). The resolution of the 3-D position measurements achievable with this particular system is 0.1 mm, assuming that the maximum distance between the transmitter (which is mounted in a belt and fixed at the back of the patient) and the sensing element mounted inside of the tip of the probe is no more than 50 cm. The pressure sensing array comprises six bending beam force sensors mounted inside the tip of the probe. The sensing part of each bending beam protrudes through the hole in the tip of the probe and
has force measuring area of 2 mm in diameter. When the tip of the probe touches the tissue, beams bend, which is detected by a pair of strain gauges bonded on opposite sides of each beam. Typical range of forces measured by each sensor is from 1 to 15 g with an accuracy of 0.2 g. The probe comprises also the ‘End of task’ button to communicate with the processor and a button for changing the position of the articulated tip of the probe. The tip of the probe is mated to a rigid tube, which in turn is attached to a pistol grip handle. A disposable rubber sheath covers the entire tip as well as the tube. Fig. 12 is a schematic diagram showing the general concept of the device and its function. The examination is conducted in accordance with a special protocol, describing a predetermined sequence of steps necessary to collect sufficient data. The computer provides feedback information on the stage and completeness of the examination process. Examination of the patient starts with inserting the probe into the rectum to the neighborhood of the prostate. At the moment when pressure sensors are passing the anus,
206
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 11. General view of the transrectal MI probe.
the processor automatically starts recording position and pressure data from the probe. The computer displays an image of typical pelvic anatomy and the position of the probe
Fig. 12. A schematic diagram showing general concept of the device and its function.
in relation to the anatomy along with a designation of the region to be scanned with the probe. This initial image serves as a map to navigate the probe. A cursor on the screen represents the tip of the examining probe. At this stage the relative positions of the probe and the prostate shown on the screen are a rough guess based on the ‘average’ male anatomy. This stage of examination is called ‘Positioning’. During this stage the operator moves the probe over the marked region to be scanned on the navigation map. The pressure exerted by the operator and the rate of movement of the probe is indicated on the screen to direct and optimize operator performance. The color of the cursor changes to indicate whether the pressure being applied is insufficient, excessive or adequate to optimally image the prostate. If the probe is being moved too fast, a warning signal is placed above the image on the screen. Similarly, if the probe is misoriented so that less than all the sensors contact tissue, an appropriate misorientation message is displayed. The aim of the ‘Positioning’ stage is to determine the exact location of the prostate in the test subject using accumulated position and pressure data and to produce an image of the prostate in the correct relation to the probe. At the end of the ‘Positioning’ stage the operator depresses a switch on the probe control handle sending an ‘End of task’ signal to the processor. Upon receiving the sig-
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
207
Fig. 13. Data collection phase of the prostate MI examination.
nal, the processor calculates the position of the center and boundaries of the prostate, thereby permitting the image on the screen to correctly show the relative position of the probe with respect to the actual prostate. These calculations take about 10–15 s, after which the image on the screen zooms out to reveal the prostate and neighboring structures. After this enlarged image of the prostate appears on the screen the process moves into the next phase. The processor program displays a sequential list of regions of the prostate to be examined adjacent to the image screen (Fig. 13) and the operator performs this sequence of operations. Advantageously, the choice of this sequence is similar to that of a regular digital rectal examination so that it is familiar to the operator. After completion of each
step of the sequence, the operator sends an ‘End of Task’ signal and the processor can check the task on the display and highlight the next task in the sequence. This data collection takes 2–3 min, after which the operator can terminate the examination by pressing the ‘End of task’ button and removing the probe from the patient. In the final step, the computer calculates a virtual pressure pattern, then based on this pattern, calculates a mechanical model of the prostate and finally, generates a three dimensional image of the examined prostate. It displays the examined prostate and in parallel, displays a reference ‘normal’ prostate. The operator is able to rotate the image of the reconstructed prostate on the screen, by using an associated computer mouse. The computer can synchronously rotate the refer-
208
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
ence prostate to facilitate comparison and detection of abnormalities. Fig. 14 presents a flow chart for data acquisition and processing in the device. The set of real time force data obtained from each sensor is processed into a set of transformed data. By elementary geometrical processing, the actual position data permits determination of three spatial coordinates for each pressure sensor. Thus, one can relate every measured force value from each sensor to a point in three-dimensional space. The force
Fig. 14. A chart showing the flow of data in the prostate imaging process.
measured by the sensor and its three coordinates are the transformed data. The transformed data can now be processed to provide pressure field data. The transformed data contain both useful information about parameters of the tissue being investigated and noise of various origins. The noise originates from force and position measurement error and from artifacts related to tissue movement (movement of the prostate, movement of the patient). Pressure field data is calculated by processing the transformed data to minimize noise and extract the 3-D spatial distribution of pressure approximating ideal conditions of measurement. While there are a number of possible algorithms for this processing, our initial approach is to use Chebyshev approximation [18]. The pressure field is represented as a superposition of Chebyshev polynomial functions. From the pressure field data approximated by Chebyshev polynomials, one can now reconstruct an elasticity model of the prostate —a model that will show the distribution of hardness in three-dimensional space. The surface of the examined prostate can be obtained by choosing a level of force corresponding to deformation of the rectal wall that permits the sensors to press against the prostate surface. From the pressure gradients, information on the prostate tissue hardness can be generated. One of the key tasks of the data processing in the MI method is to relate the data on the spatial distribution of the elasticity of the prostate to a database of general and patient specific medical data. In traditional medical imaging, the device usually displays the structure of an object in terms of some measured physical property. The image obtained this way is often very far from what the actual examined region of body or an organ would look like if exposed to direct sunlight, or
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
drawn by an artist. Therefore, an expert in a particular type of image analysis is required to tell the physician what information from the image is relevant to the diagnosis. Currently, as a result of a wider use of powerful computer means and databases, an alternative approach to imaging, so called knowledge-based imaging has started to emerge. Using knowledge-based imaging, a computer can store in memory a 3-D picture of a ‘normal’ prostate which is being examined and adjust (transform) this image according to the measured data, to produce an image that represents the actual examined gland. Such a pictorial 3-D image or its cross sections will additionally include data on the mechanical properties of the prostate. It will be significantly easier for a physician to recognize abnormalities of the examined organ, represented on such an image. Further, the expert system will use the knowledge about characteristics of different types and stages of prostate cancer to point out any poorly defined and suspicious regions in the model, or just show any abnormalities or deviations from what the ‘normal’ prostate should look like. At this point, the physician can also enter into the computer new information based on other tests or exams performed on the same prostate and the knowledge base will ‘learn’ and ‘expand’. Once a 3-D model of the actual examined prostate is stored in the computer, it must be presented by the user in a way that would allow both external and internal features to be seen on one picture. This means that the 3-D image on the screen should contain information about geometrical features of the prostate as well as spatial distribution of elasticity and surface texture information. Additionally, the image should indicate to the user which areas of the examined prostate are poorly defined and need to be examined further in order to produce a complete diag-
209
nosis. There are several potential 3-D visualization methods that can be suitable for this task, such as polygon based surface methods, ray cast volume rendering and cross section slicing. In order to utilize knowledge-based imaging methods in our project, we need to have a 3-D model of prostate in the computer which can be easily redimensioned to fit our measured data. For the internal computer representation of such modeled surfaces of the prostate, parametric spline surfaces (X(s, t), Y(s, t), Z(s, t)) comprised of triangular surface patches which are spliced together so that we have an overall smooth surface can be used. Each of these surfaces has a fixed topology in that the number and connectivity of individual triangular patches does not change but each has control point parameters which are used to control the shape and geometry of the surface. The actual values of the control point parameters are eventually selected so as to ‘best fit’ the data of a particular gland, but the number of surfaces and the number of triangular patches comprising each surface remains fixed (unless the user intervenes and requests a change).
3.1. Software for the prostate MI de6ice The software of the device is implemented as a 32-bit application for Microsoft ® Windows NT™ operation system. The user interface module is developed as a main EXE module and all calculation modules are implemented as a Windows function in a DLL. For the user interface module development we used the visual programming environment for rapid application development (RAD) Delphi 2.0™ and the program language Object Pascal. The user interface can provide two windows: the main window consisting of control elements which allow the user to
210
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
manage the process of examination and the visualization window for 3-D or 2-D representation of the imaged prostate. Key functions of the software are numerical filtration of the pressure and coordinate data flow and reconstruction of the 3-D virtual pressure pattern characterizing both geometrical and mechanical structure of the examined prostate. Reconstruction of the 3D virtual pressure pattern is made with the use of Chebyshev polynomials. The approximated function is given as a linear superposition of the basis functions where products of Chebyshev polynomials from spatial coordinates are used. The approximation parameters are the coefficients in such a linear combination. The coefficients are determined as a solution of a linear system of equations. The computer has in memory a 3-D model of a ‘normal’ prostate and adjusts (transforms) this model according to the measured data to produce an image that represents the actual examined organ of a particular patient. A number of potential visualization methods capable of displaying a realistic 3-D image of a prostate on the screen were evaluated. The decisions on the adequacy of a method for our objectives were based on its ability to not only display the geometrical features of the prostate, but also to indicate the spatial hardness and texture distribution in a clear and easily recognizable way. The tasks on pictorial representation of structural and mechanical features of the prostate were accomplished and a 3-D computer model of normal and diseased prostates were developed in collaboration with Dr. Marsha Jessup, Director of the Media Resources Department, Robert Wood Johnson Medical School, New Jersey. Fig. 15 shows a 3-D model of a normal prostate and a series of 3-D models of a prostate with two benign tumors near the urethra and a growing car-
Fig. 15. Three-dimensional computer model of a normal prostate and a series of 3-D models of a prostate with two benign tumors near the urethra and a growing carcinoma in the outer area of the prostate.
cinoma in the outer area of the prostate. In addition, 3-D computer models of the male urogenital system and rectum anatomy have been developed (Fig. 16) as a part of the anatomical database which will be needed in the specialized software and interactive diagnostic algorithms in the knowledge-based mechanical imaging device.
Fig. 16. Three-dimensional computer model of the male urogenital system and rectum anatomy.
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
The 3-D prostate reconstruction from the experimental data is based on the Finite Elements Method (FEM) widespread in CAD/ CAM systems to provide maximum reality in the displayed model. The model includes information on geometry as well as information on the tissue elasticity and possible nodules and is based on the use of isoparametric 3-D elements having 21 nodes and shelled 3-D isoparametric elements using eight nodes. A special fast algorithm of photorealistic image reconstruction has been developed. Brightness of different sections of the image surface depends on their inclination and takes into consideration both scattered and mirror-image components of reflection. The resources of the True Color palette are fully used. Using knowledge-based imaging methods requires having a 3-D model of a prostate in the computer that can be easily redimensioned to fit measured data. For the internal computer representation of such modeled surfaces of the prostate, parametric spline surfaces (X(s, t), Y(s, t), Z(s, t)) comprised of triangular surface patches which are spliced together to create an overall smooth surface can be used. Each of these surfaces has a fixed topology in that the number and connectivity of individual triangular patches does not change, but each has control point parameters which are used to control the shape and geometry of the surface [19]. The actual values of the control point parameters are eventually selected so as to ‘best fit’ the data of a particular gland, but the number of surfaces and the number of triangular patches comprising each surface remains fixed. To analyze quantitatively the spatial resolution of the device, to express accuracy and tolerance in millimeters and to be able to compare performance of the MI method with
211
that of DRE, a pilot experimental study has been performed. Fig. 17 shows cross-sections of the measured (upper row) and reconstructed (lower row) model prostates. Solid lines in the upper row figures represent actual profiles of the examined models and experimental points shown in the same figures represent filtered data. Fig. 17A shows the data obtained by sensing a prostate phantom by the MI probe directly and the data of Fig. 17B represents a similar experiment on the same model with one difference: the phantom was covered by a 3 mm thick rubber layer simulating the rectal wall. These experimental studies demonstrated the following capabilities of the developed software: the influence of the ‘rectal wall’ on the evaluation of the prostate parameters is not significant; the data acquisition part of the device is capable of reproducing geometrical features of the prostate with tolerance in the 1–2 mm range; knowledge-based reconstruction algorithms do not add significant distortions in the simulated prostate 3-D model. Special experiments were performed to obtain a quantitative comparison of the MI and DRE evaluation of the geometrical features of the prostate models. A number of researchers and doctors were asked to palpate different prostate models through an intermediate layer of soft rubber and record their observations. Specific parameters including width and length of the model, depth of median groove and asymmetry of the model were noted. The MI device was then used to examine the models. The MI device was found to evaluate all geometrical parameters about five times more accurately than the human finger. This means that in a simplified model situation were only geometrical features are considered, we can potentially have
212
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 17. Cross-sections of the measured and reconstructedmodel prostate: (A) without intermediate layer; and (B) covered by a 3 mm thick rubber layer.
five times improvement in examination accuracy when MI is used compared to DRE. Experiments on the prostate phantoms were performed to test developed algorithms of knowledge-based 3-D reconstruction of the prostate (Figs. 18 and 19). Fig. 18 shows experimental data on the ‘equal pressure’ patterns (see explanations for Fig. 10) obtained for a prostate phantom shown schematically on the right panel of the figure. The prostate phantom during the measurements was covered by 3 mm thick layer of soft rubber simulating the rectal wall. Fig. 19A shows calculated surface of the model using the ‘equal pressure’ patterns based on the data received from the force and position sensors. The 3-D pattern obtained enables quantitative evaluation of the geometrical features of the examined model such as dimensions, symmetry, presence and depth of median groove, etc. Fig. 19B is a result of knowledge based
3-D reconstruction of the prostate using the data of Fig. 19A. These experiments have shown that all the important and diagnostically relevant features of the prostate model are reconstructed without significant distortion.
4. Conclusions Results of the laboratory pilot studies have proven the feasibility of the mechanical imaging technology. The project on development of the first MI device for diagnosing prostate diseases and detecting prostate cancer has now moved into the phase of clinical trials. The Robert Wood Johnson Medical Center Institutional Review Board has approved the clinical testing of the prototype and a number patients were examined and 3-D mechanical images of in vivo prostates have been ob-
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
213
Fig. 18. Experimental data on the ‘equal pressure patterns’ for a prostate phantom shown schematically on the right panel of the figure.
tained. Clinical studies were conducted under the guidance of Professor K.B. Cummings and Dr R.E. Weiss. Here we present just two examples of the clinical examination data. A complete report on the results of the clinical studies will be presented elsewhere. Fig. 20 shows comparative projections of three-dimensional images of the first mechanically imaged prostate of a
healthy volunteer (right column) and a computer simulated ‘ideal’ prostate (left column). There were no abnormalities seen and the prostate appeared to have a uniform consistency. Fig. 21 presents the first MI data on a patient with histologically confirmed prostate cancer. This patient was noted to have a PSA of 11.1 (normal 1–4.0) and had a small nod-
214
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
Fig. 19. Knowledge based three-dimensional reconstruction of the prostate phantom.
ule noted on digital rectal examination. Because of these findings, rectal sonography and random prostate needle biopsies were performed. The rectal sonography revealed an enlarged 58.8 g. prostate, which had no
ultrasonically visible abnormalities (upper right panel in Fig. 21). Random biopsies revealed a 2/5 Gleason score prostate cancer on the left side of the prostate, which correlates with the nodule palpated on DRE. MI data shown in Fig. 21 as 3-D virtual strain pattern (upper left panel) and a cross-section of the force/position data across the prostate apex (lower panel), clearly show the location and dimensions of the prostate cancer nodule missed by ultrasound. These promising results portend the creation of a simple and inexpensive device for imaging the prostate, evaluating its state and diagnosing prostate diseases, which utilizes physical principles and measured parameters similar to those associated with a digital rectal examination. Preliminary data obtained in the pre-clinical testing of the prostate MI device strongly suggest that mechanical imaging technology meets basic requirements for a mass cancer screening method and for an inexpensive and affordable method of day-to-day monitoring of cancer in its advanced stages: it is simple, fast, inexpensive and safe.
Fig. 20. Comparative projections of three-dimensional images of the first mechanically imaged prostate of a healthy volunteer (right column) and a computer simulated ‘ideal’ prostate (left column).
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
215
Fig. 21. Comparison of the MI and sonography data obtained for a patient with a palpable nodule in the left part of the prostate apex where the prostate cancer was revealed by random biopsy.
However, the MI method is general and has much broader implications in a variety of diagnostic and surgical applications that require cancer diagnosis and assessment and localization of abnormal tissue for biopsy and characterization. Acknowledgements This work was supported in part by the National Cancer Institute (USA) grant 1 R43 CA69175-01. References [1] A.P. Sarvazyan, A.R. Skovoroda, Method and apparatus for elasticity imaging, US Patent 5524636, 1996.
[2] A.P. Sarvazyan, Knowledge-based mechanical imaging. Proc. 10th IEEE Symp. on Computer Based Medical Systems’97, IEEE Computer Society publication, Washington, 1997, pp. 120 – 125. [3] P. Strax, Control of breast cancer through mass screening, Hospimedica March/April (1989) 35–40. [4] C.C. Bohring, T.S. Squires, T. Tong, Cancer statistic, CA Cancer J. Clin. 43 (1993) 7 – 26. [5] W.J. Catalona, J.P. Richie, F.R. Ahmann, et al., A multicenter evaluation of PSA and digital rectal examination (DRE) for early detection of prostate cancer in 6374 volunteers, J. Urol. 149 (Suppl) (1993) 412A. [6] D.S. Smith, W.J. Catalona, Inter-examiner variability of digital rectal examination in detecting prostate cancer, Urology 45 (1995) 70 – 74. [7] P.J. Littrup, A.C. Goodman, C.J. Mettlin, The benefit and cost of prostate cancer early detection, Cancer J. Clin. 43 (1993) 134 – 149. [8] E.H. Frei, B.D. Sollish, S. Yerushalmi, Instrument for viscoelastic measurement, U.S. Patent c 4,250,894, 1981.
216
A. Sar6azyan / International Journal of Medical Informatics 49 (1998) 195–216
[9] C.R. Gentle, Mammobarography: a possible method of mass breast screening, J. Biomed. Eng. 10 (1988) 124–126. [10] A.P. Sarvazyan, A.R. Skovoroda, S.Y. Emelianov, J.B. Fowlkes, J.G. Pipe, R.S. Adler, R.B. Buxton, P.L. Carson, Biophysical bases of elasticity imaging, in: J.P. Jones (Ed.), Acoust. Imaging, vol. 21, Plenum, New York, 1995, pp. 223 – 240. [11] J.N. Goodier, Concentration of stress around spherical and cylindrical inclusions and flaws, Trans. ASME 55 (39) (1933) 39–44. [12] A.P. Sarvazyan, F.L. Lizzi, P.N.T. Wells, A new philosophy of medical imaging, Med. Hypotheses 36 (1991) 327–335. [13] Y.P. Pyt’ev, Measurement computer systems of superhigh resolution, Pattern Recognit. Image Anal. 1 (1) (1991) 54–76. [14] Y.P. Pyt’ev, Fuzzy uncertain computer-aided measuring systems, Pattern Recognit. Image Anal. 4 (1994) 178–184.
.
[15] D.A. Pashko, Y.P. Pyt’ev, A.P. Sarvazyan, A.R. Skovoroda, On the ultimate potentialities of classification of pathologies in the problem of diagnostics of breast cancer, Pattern Recognit. Image Anal. 6 (1996a) 510 – 522. [16] D.A. Pashko, Y.P. Pyt’ev, A.P. Sarvazyan, Minimax evaluation of parameters of a nodule in diagnosing breast cancer with the use of a force sensor array, Bulletin of Moscow State University, Ser. 3, Physics, Astronomy (1996b) 18 – 25. [17] A.R. Skovoroda, Inverse problems of the theory of elasticity in the diagnostics of soft tissue pathologies, Preprint, Pushchino Research Center of the Russian Academy of Sciences, Pushchino (1992). [18] J.P. Boyd, Chebyshev and Fourier Spectral Methods, Springer, New York, 1989. [19] W. Welch, A. Witkin, Free-form shape design using triangulated surfaces, Comput. Graph. 28 (1994) 247 – 256.