- Email: [email protected]
A novel breast ultrasound system for providing coronal images: System development and feasibility study
Ultrasonics 56 (2015) 427–434
Contents lists available at ScienceDirect
Ultrasonics journal homepage: www.elsevier.com/locate/ultras
A novel breast ultrasound system for providing coronal images: System development and feasibility study Wei-wei Jiang a, Cheng Li b,c, An-hua Li b, Yong-Ping Zheng a,⇑ a
Interdisciplinary Division of Biomedical Engineering, The Hong Kong Polytechnic University, Kowloon, Hong Kong, China Department of Ultrasound, State Key Laboratory of Oncology in Southern China, Sun Yat-Sen University Cancer Center, Guangzhou, China c Department of Ultrasound, Hospital of Traditional Chinese Medicine of Zhongshan, Zhongshan, China b
a r t i c l e
i n f o
Article history: Received 9 May 2014 Received in revised form 10 September 2014 Accepted 16 September 2014 Available online 27 September 2014 Keywords: Coronal image Breast ultrasound Image rendering Breast cancer Breast diagnosis
a b s t r a c t Breast ultrasound images along coronal plane contain important diagnosis information. However, conventional clinical 2D ultrasound cannot provide such images. In order to solve this problem, we developed a novel ultrasound system aimed at providing breast coronal images. In this system, a spatial sensor was fixed on an ultrasound probe to obtain the image spatial data. A narrow-band rendering method was used to form coronal images based on B-mode images and their corresponding spatial data. Software was developed for data acquisition, processing, rendering and visualization. In phantom experiments, 20 inclusions with different size (5–20 mm) were measured using this new system. The results obtained by the new method well correlated with those measured by a micrometer (y = 1.0147x, R2 = 0.9927). The phantom tests also showed that this system had excellent intra- and inter-operator repeatability (ICC > 0.995). Three subjects with breast lesions were scanned in vivo using this new system and a commercially available three-dimensional (3D) probe. The average scanning times for the two systems were 64 s and 74 s, respectively. The results revealed that this new method required shorter scanning time. The tumor sizes measured on the coronal plane provided by the new method were smaller by 5.6–11.9% in comparison with the results of the 3D probe. The phantom tests and preliminary subject tests indicated the feasibility of this system for clinical applications by providing additional information for clinical breast ultrasound diagnosis. Ó 2014 Elsevier B.V. All rights reserved.
1. Introduction Breast cancer is the most common cancer in women worldwide. In the Global Health Estimates of World Health Organization (WHO), it was estimated that 508,482 women died of breast cancer in 2011 in the world [1]. In America, it was reported that 226,870 women were diagnosed with breast cancer and 39,510 of them died of breast cancer in 2012 [2]. According to the report of Breast Cancer UK, breast cancer accounted for 31% of cancers diagnosed in women [3]. Up to now, there has not been an effective method to prevent breast cancer and early detection has remained the cornerstone for breast cancer control [4]. Among all breast cancer detection methods, ultrasound plays an important role in breast cancer deaths decline for its advantages of radiation-free, real-time and ⇑ Corresponding author at: Interdisciplinary Division of Biomedical Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, China. Tel.: +852 27667664; fax: +852 23624365. E-mail address: [email protected] (Y.-P. Zheng). http://dx.doi.org/10.1016/j.ultras.2014.09.009 0041-624X/Ó 2014 Elsevier B.V. All rights reserved.
suitable for dense breast [5,6]. Ultrasound has long been recognized as a valuable tool to distinguish between cysts and solid masses. With the rapid development of ultrasound techniques and greatly increased images quality, breast ultrasound can now not only be used for characterizing cysts, but also differentiating benign from malignant lesions. In a breast abnormalities (259 carcinomas, 1820 benign) examination, ultrasound could help to avoid unnecessary biopsy with benign diagnosis results in 71 suspicious cases at palpation or mammography [7]. Therefore, routine ultrasound examination can help to reduce unnecessary biopsies. In clinical breast ultrasound examination, 2D ultrasound probe is routinely used which can only provide transverse and longitudinal images but no coronal images. However, information on this plane has been proved to be beneficial for clinical diagnosis [8– 14]. Rotten et al. analyzed images of normal breast tissue and breast lesions and found four diagnosis features on coronal plane [8]. Among these features, one was defined as compressive pattern which was thought to be associated with benign lesions. In this pattern, the continuous hyperechoic bands of tissue peripheral to
428
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
the masses appeared to be distinct from the central part. Another feature was called converging pattern which was a typical characteristic of malignant lesions. For this pattern, a stellate distortion consisting of alternating hypoechoic and hyperechoic lines converged towards to the hypoechoic central masses [8]. In the study of Chen et al., the two features were described as hyperechoic rim and retraction phenomenon. They were used independently to differentiate breast lesions and good accuracy (95.9% for hyperechoic rim; 96.8% for retraction phenomenon) and specificity (92.8% for hyperechoic rim; 100% for retraction phenomenon) were reported [9]. In fact, the feature of retraction phenomenon on coronal plane, also called spiculation, was reported by other researchers with high specificity (98.4% in [10]; 94.6% in [11]) in tumor diagnosis. In the report of Meyberg-Solomayer et al., the coronal plane could provide tumor classification information when the infiltrative zone was not visible (17 of 39 cases) or unclear (6 of 8 cases) in 2D ultrasound imaging [12]. This result demonstrated that the image on this plane could offer a better assessment when the infiltrative zone surrounding the lesion was unclear or not visible on conventional 2D images, which could help to reduce biopsies. Images along coronal plane were also beneficial for tumor extent measurement [13] and ultrasound-guided vacuum-assisted core-needle biopsy [14]. Based on images on coronal images, various computer-aided diagnosis (CAD) methods were presented to help to automatically detect tumor candidate [15], mark spiculated masses [16,17], and classify tumor stages [18,19]. Breast coronal images can be provided by the technique of three-dimensional (3D) ultrasound imaging. Many researchers have been studying on this technique. One approach of 3D breast ultrasound imaging was to scan the breast using the conventional 2D probe, which was driven by a mechanical motor [20–22]. A typical representation of this approach was the method proposed by Kotsianos-Hermle et al. [20]. In this method, breast was compressed by two paddles and a probe was driven mechanically on the top of the paddle to acquire the breast images. Another approach was to scan the patient with a specially designed probe when the patient was in supine position, such as the commercial products Automated Breast Volume Scanner (ABVS) of Siemens and the Automated Whole-Breast Ultrasound (AWBU) of Sonocine [23–25]. Another approach for 3D ultrasound imaging is the 2D array ultrasound probe which uses electronic pyramidal scanning [26]. They have been used successfully for real-time 3D imaging of the heart, where high volume frame rate (40 volumes per second) is required [27], but are seldom used for breast imaging. A high volume frame is not as necessary for breast imaging, unless breast tissue motion needs to be tracked in three dimensions in real time, as in 3D elastography [28]. Therefore, the first two approaches of 3D ultrasound imaging can provide breast coronal images. However, in these methods, the driven motor or specially designed probes were required for scanning. These equipments were bulky and large, which were inconvenient for clinical scanning to offer regular motions. In addition, the moving manner of the probe in these systems was predefined so the operator could not move the probe to the desired position freely. Some regions such as axillary region and tissue against the chest wall were not accessible by using these systems. Therefore, there are still many works to be done before 3D ultrasound imaging technique can be widely used in clinical breast examination. 2D ultrasound imaging remains the dominant scanning mode for clinical breast ultrasound diagnosis. Accordingly, this study was aimed to develop a breast ultrasound system for providing coronal images based on the clinical 2D ultrasound scanner. In the following sections, this system is described in details. The system accuracy and reliability tests based on phantoms are presented. Preliminary clinical tests were also performed to demonstrate the system feasibility.
2. Methods 2.1. System overview A corresponding freehand 3D ultrasound annotation system was previously developed and successfully used for annotating breast ultrasound images [29]. Fig. 1 shows the diagram of this system. It consisted of three main components: an ultrasound machine (EUB-8500, Hitachi, Tokyo, Japan) with a linear 2D probe (EUP-L65/6–14 MHz, Hitachi, focused probe, 6–14 MHz), an electromagnetic spatial sensing device (med-SAFE, Ascension Technology, Burlington, VT, USA) and a computer with Intel Core i5 3.35 GHz CPU and 3.5 GB of memories. A video capture card (NIIMAQ PCI/PXI-1411, National Instruments Corporation, Austin, TX, USA) and a customized program were installed on this computer. The electromagnetic spatial sensing device was employed to acquire the image spatial data in this system. The selected device had high spatial accuracy. The documented positional accuracy and angular accuracy of this device were 1.4 mm and 0.5°, respectively (medSAFE Manual, Ascension Technology). The spatial device was comprised of a control box, transmitter and a sensor. The diameter of the cylindric sensor was 2.0 mm and the length was 9.9 mm. This sensor dimension was small so it was easy to be fixed on the ultrasound probe by a custom-designed kit. The image spatial data acquired by this sensor included three positions (x, y, z) and three orientations (azimuth, elevation, roll). These data were sent from the control box of the spatial device to the computer through its serial port. The sampling rate of medSAFE was 100 Hz, which was higher than the ultrasound imaging rate. So sufficient spatial data were collected and averaging was used to improve the accuracy of the system on distance and angle measurement. During scanning, the video stream of 2D B-mode ultrasound images was captured by the video capture card and sent to the computer. Meanwhile, the spatial data of these images were also sent to the computer by the control box of the spatial device. The developed program acquired and recorded these images together with their corresponding spatial information for the further visualization and rendering. Spatial calibration experiments were performed for the system to determine the position and orientation offsets between the ultrasound image and the spatial sensor. A cross-wire phantom was used to calibrate this system [30,31]. Two wires were crossed and submerged in a water tank and the wire ends were fixed to the tank. The ultrasound probe was moved slowly to scan the wire cross. If an image with a clear cross was found, this image and its spatial information would be recorded. In each experiment, 60 images from various directions were captured. According to the
Fig. 1. Diagram of the breast ultrasound rendering system developed in this study, which consisted of an ultrasound machine, a spatial sensing device and a computer.
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
pixel position of the cross on each image and the positional data read from the spatial sensor, the spatial transformation was then calculated using the Levenberg–Marquardt nonlinear algorithm [32]. Levenberg–Marquardt nonlinear algorithm was a common technique to solve nonlinear least squares problems. This technique took advantages of both the gradient descent and Gauss– Newton methods. It acted more like a gradient-descent method when the parameters were far from their optimal values, and acted more like the Gauss–Newton method when the parameters were close to their optimal values. This method was robust and popular in solving nonlinear least squares problems so it was used in this study. 2.2. Data acquisition and volume rendering A program for this system was developed using Visual C++ (Microsoft, Redmond, WA, USA) and Visualization Toolkits (VTk, Kitware Inc., NY, USA). The software interface is shown in Fig. 2, which consisted of control bar and the display window. The main functions on the control bar could be organized into four parts: data acquisition, data processing, image rendering and visualization. The data acquisition part was designed for the capture controlling of ultrasound images and spatial data. The acquisition was in real-time mode. Acquisition parameters such as image size could be set by the operator in this part. The processing part contained functions to manually choose valid image frames and to select region of interest (ROI) after data were captured. The processing functions were conducted in off-line mode. The rendering function was responsible for the volume rendering to obtain the coronal images. For this system, the off-line rendering procedure normally took less than 10 s. After rendering, the coronal image was displayed on the display window. This procedure was finished by the visualization function, which took less than 1 s. The visual-
429
ization function also included the real-time display of the original ultrasound images during the acquisition. During the scanning, the operator held the probe and moved freely on the subject. Ultrasound images together with their corresponding spatial data were displayed on the software interface in real time, also shown in Fig. 2. To control the image acquisition during scanning the subject by hands, a foot switch was designed. When the operator stepped on the switch, images with their corresponding spatial data were recorded. If the operator stepped again, the recording was stopped. During scanning, if features of breast tumor such as irregular border were found on images, the operator could step on the foot switch to record images and spatial data for the further processing. For different operators, the scanning speeds were different. This had influence to the reconstructed image resolution along the scanning direction as the frame rate was fixed, which was 21 frames per second for our system. When the probe was moved faster, the distance between two consecutive 2D images would be larger, thus lower resolution along the scanning direction. According to the frame rate collecting B-mode images, the probe could be moved as fast as 10 mm/s to achieve a gap smaller than 0.5 mm. Thus, using a normal scanning speed, the scanning gaps would be very small and be filled through the later interpolation. Furthermore, after scanning, the acquired images could be displayed in 3D space so that the operator could investigate if there was region not covered. If obvious gaps were found, this data sequence would be discarded and the operator may scan the lesion region again. In clinical breast ultrasound images, since the useful information such as lesion region was only small part on raw 2D ultrasound images, especially for small lesions (<10 mm diameter) [33]. In clinical breast imaging, small lesions were more easily missed by the clinicians. Therefore, in this study, the region of interest (ROI) should be defined before rendering. In the present
Fig. 2. The software interface of the breast ultrasound rendering system comprising the control panel and display window.
430
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
Fig. 3. The method used for defining region of interest for rendering.
method, the start and end B-mode image including diagnosis information were firstly marked by the operator when reviewing images, as shown in Fig. 3. To get the region of interest from original images, the operator could set two points on one image to define a region including the lesion. On the basis of the two points, the program could automatically calculate the region of interest and discard the information outside this region, as shown in Fig. 3. The conventional volume rendering methods usually contained two steps. The first step was to reconstruct a regular voxel array using the recorded raw images. Then the voxel array was visualized using the planar reslicing. The two steps involve two resampling stages, from raw image pixels to the voxel array, and from the voxel array to the slice pixels. However, for the coronal imaging purpose of this study, the first step to form the voxel array was not necessary. In this study, a volume rendering approach, which obtained the reslice plane with a certain thickness directly from the raw data, was used. This approach only required one resampling stage. Since the resampling procedure inevitably involves the data approximation, this approach was expected to provide more accurate visualization. This rendering method, also called narrow-band volume rendering, was first introduced by Gee et al. [34–36]. Since the coronal image was endowed with a certain thickness based on the raw image data, the narrow-band meant a layer with limited thickness and the value of thickness could be determined by the operator. In this study, the thickness ranged from 4.8 mm to 9.7 mm. In this volume rendering method, as shown in Fig. 4, all selected 2D images were projected to the coronal plane to form a new 2D coronal image. First, according to the
Fig. 4. An illustration of the volume rendering method in this developed system.
scanning region and the size of ROI, the coronal image coordinate system with a regular pixel array was defined. The density of the matrix was determined by operators during reconstruction, while the resolution of 2D imaging plane was determined by the ultrasound transducer. On the basis of the spatial data and the calibration matrix, each pixel on the original 2D images was transformed to the new coordinate system. The final value of the pixel intensity on the new coronal image was calculated based on all pixels falling into its region. Two methods including minimum intensity compounding and averaging were used to calculate the intensity. The minimum intensity compounding was to choose the minimum intensity from all pixels falling into the pixel region as the final value of this pixel. The average method was to average all pixels in the region and the average value was used as the final value. Minimum intensity compounding was suitable for the fluid-filled lesion such as cyst. Simple averaging could produce an X-ray like image [36]. After the data mapping, the intensity of most pixels could be calculated. However, in this method, the free-hand scanning was adopted. This scanning method might cause gaps though it had special advantages such as the flexibility of manipulation [27,37]. In order to fill gaps, interpolation was used to compute all empty grids. A variety of methods, such as the nearest neighbor interpolation and linear interpolation, had been reported for the gap filling [38–40]. In this system, the bilinear interpolation was used to produce the trade-off between the interpolation performance and computation time. For each empty pixel, its intensity was the average of non-empty pixels in the nearest 2-by-2 neighborhood. 2.3. Validation phantom and experiments To demonstrate the system accuracy and repeatability in providing coronal images, phantoms were designed for validation experiments. Today, various commercial tissue phantoms are available for different medical applications. However, in this study, phantoms containing 20 sphere inclusions with different dimensions were needed. There are no commercial phantoms available with all these different sizes of inclusions. Therefore, the customdesigned phantoms were employed in this study. The diagram of the phantom was shown in Fig. 5(a), which comprised of matrix and inclusion. In order to mimic the clinical ultrasound images of breast tumor, the material of matrix should be hyperecho and the inclusion should be hypoecho. The inclusion was made from a material called crystal soil which was a water-absorbent polymer. The crystal soil could hold up to 180 times it volume in water therefore the main component of the inclusion was water, as shown in Fig. 5(b). Fig. 5(c) is the matrix which was a mixture of water, 5% gelatin (Gelatin from porcine skin, Type A, Sigma Aldrich, St. Louis, MO, USA) and 3% agar (Agar, granular powder, general purpose grade, Fisher Scientific, Loughborough, UK). The agar was used as the scatter to make the matrix in hyperecho status. The mixture was initially heated to 50 °C. This temperature could keep the mixture in the liquid status and the agar was not dissolved so the agar particles could provide speckles. And then the temperature was reduced to 24 °C quickly to avoid agar sedimentation because the mixture began to solidify at 24 °C [41,42]. Before the matrix was totally solidified, the inclusion was put into the matrix using tweezers. And then the mixture with inclusion was put into the refrigerator for 12 h for the further solidification. The ultrasound image of the phantom is shown in Fig. 5(d), which was similar to the clinical images of breast cancer, especially the cyst cases. In validation experiments, the accuracy and repeatability of this system were tested. For accuracy test, 20 inclusions with different size (5–20 mm) were used. The diameters of these inclusions were firstly measured by a micrometer. For each inclusion, the measure-
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
431
Fig. 5. The validation phantom. (a) The diagram of validation phantom; (b) the inclusion of crystal soil; (c) the phantom matrix; (d) the ultrasound image of the phantom.
ment was repeated three times and the average was used as the true size of the inclusion. Then these inclusions were put into matrix. Totally five phantoms were made and each of them included four inclusions. These phantoms were scanned by the system and coronal images were obtained after volume rendering. The inclusion dimension was then measured on the coronal images, here the coronal plane refers to the plane parallel to the phantom surface, which are similar to the coronal images in breast ultrasound imaging. The measurement results were compared with the real size of these inclusions to test the system accuracy. For reliability tests, the operator held the probe to scan the phantom and then coronal images were reconstructed to measure the inclusions size. In intra-rater tests, one operator scanned these phantoms twice at the intervals of at least 5 h. For inter-rater tests, the other operator performed the same procedure on phantoms after the first operator finished. The parameter of Intraclass Correlation Coefficient (ICC) was used to analyze data [43,44]. The intrarater and inter-rater repeatability was, respectively, assessed by the ICC (2, 1) and ICC (3, 1) [45]. All the statistical analyses were conducted by the statistical software SPSS (SPSS Inc., Chicago, USA). 2.4. In vivo experiments Although phantom experiments could provide evidence for the feasibility of this new method, in vivo experiments were more important to demonstrate the system performance on the clinical application. In the experiments, a commercial ultrasound machine (Logiq E9, GE Healthcare, New York, USA) with a linear 2D probe (ML6-15-D, GE Healthcare, New York, USA) and a 3D probe (RSP6-16-D, GE Healthcare, New York, USA) was employed to test the feasibility of the new system. Since the ultrasound scanner used in our system did not have a 3D probe, the above commercial scanner which had a 3D probe was used in the in vivo experiments. The aim of the subject tests was to compare the performance of our system with that of a commercial 3D probe. This 3D probe was a routinely used in clinics so it met our requirement. Three patients (women, 63, 57, and 53 years old), who had breast pain or findings using palpation, were recruited to have ultrasonography. After ultrasound examinations, the three patients had biopsies. The histopathological results showed that they were all ductal carcinoma.
During the scanning by our system, the patient lay on the bed in supine position. An operator held the probe to scan the breast in the orthogonal antiradial scanning pattern [46,47]. When a lesion was found, the operator would not stop scanning but remember the lesion location until covering two whole breasts. This step could help to ensure that all breast regions were covered. After two breasts were scanned, the operator moved the probe to the lesion region to scan this part slowly and saved images together with their corresponding annotation information. The above scanning procedure was the routine clinical breast ultrasound examination procedure. The total scanning time, including the time spending on scanning two breasts and scanning the lesion region, was recorded. After the scanning using the new system, the patient kept the same posture on the bed and was scanned using a combination of the linear probe and the GE mechanically scanning 3D probe. The operator held the linear probe and scanned the two breasts to find the lesion region. Then she changed to the 3D probe and put the probe on the lesion part. The 3D probe was kept stationary on the lesion for about 15 s until the acquisition was finished. The scanning time for this procedure was also recorded. The in vivo experiments were conducted in the Sun Yat-Sen University Cancer Center (Guangzhou, China). The patients were given the informed consent before scanning, which was approved by the human subject ethics committee of the institution. To demonstrate the performance of the new system in clinics, the required scanning time and the formed coronal images of the two methods were compared. In this study, only the scanning time, which was spent on acquiring images, was compared. The time spending on the coronal image reconstruction was not included because it was in the review section in clinical breast examination. The review section was off-line. Therefore, compared with the time spending on review section, the scanning time was more important. The time saved during scanning section could help to scan more patients in clinical examinations. For the commercially available GE 3D probe, after image reconstruction, slices along the coronal plane could be provided for different depths. After reviewing all slices, the plane with largest dimension in tumor size was selected and the width and height of the tumor were measured. The depth of the selected slice was also recorded. Using the new system, the coronal image at the same depth was reconstructed. The tumor
432
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
size was also measured on this image and compared with the GE 3D scanner results.
Table 1 Comparison of scanning time of the two methods.
3. Results 3.1. Validation results
By 3D probe By new system
Patient 1 (s)
Patient 2 (s)
Patient 3 (s)
Average time (s)
75 62
73 66
75 63
74.3 63.7
Fig. 6 is a typical coronal image of the phantom. The dark part indicated the inclusion. The dimensional measurement was also shown on the image. The measurement results were compared with the actual sizes of inclusions to demonstrate the system accuracy. Fig. 7 illustrates the correlation about the phantom diameters measured by the 3D ultrasound rendering image and the micrometer. The results demonstrated a very good linear correlation (y = 1.0147x, R2 = 0.9927) between the results obtained by the two methods. This indicated that the measurement using the new system well matched the actual values. The ICC values for intra- and inter-rater were 0.999 and 0.995, respectively. The high values indicate excellent repeatability of the new system. 3.2. Results for the in vivo experiments The scanning time for the two methods is summarized in Table 1. Using the commercial linear probe together with the 3D probe, the average scanning time was 74 s. And for the new system, the mean acquisition time was 64 s. Compared with the commercial scanner, this new system required shorter scanning time.
Fig. 8. The comparison of coronal images from two methods. (a) Coronal image from GE 3D scanner; (b) coronal image obtained by our rendering system.
Table 2 The tumor size comparison between the two methods.
By 3D probe By new system Difference
Fig. 6. The typical coronal image and dimensional measurement of phantom for accuracy test, here the coronal plane refers to a plane in parallel to the phantom surface.
Patient 1 (mm)
Patient 2 (mm)
Patient 3 (mm)
Width
Height
Width
Height
Width
Height
18.9
19.3
23.3
17.7
18.0
19.2
17.7
17.3
21.7
15.6
17.0
17.2
1.2 (5.7%)
2.0 (10.4%)
1.6 (6.9%)
2.1 (11.9%)
1.0 (5.6%)
2.0 (10.4%)
The reason for the longer time of the commercial scanner was probably due to the switch between probes. In clinical breast examination, two steps are usually needed. The first step is scanning the whole breast to find all lesions. For this step, the linear probe ML6-15-D was used and it could allow the operator to investigate tumor in real-time on the screen. Then the operator put down the linear probe and switched to the 3D probe to image the tumor region. The probe switching obviously required additional scanning time and caused inconvenience for the examination. Fig. 8(a) shows the coronal image obtained using the GE 3D probe. As shown in the figure, the clinician manually put cross on the image to measure the tumor width and height. After setting depth and thickness, the coronal image on the same position could be provided by the rendering method of the new system, as shown in Fig. 8(b). The tumor size could also be measured by the operator on this image. The comparison of tumor size between the two methods for three patients is summarized in Table 2. As shown in Fig. 8, the tumor size is represented by width ⁄ height in mm. For the three lesions, the differences between the results obtained by the new method and the GE scanner were 1.2 ⁄ 2.0 mm (5.7% ⁄ 10.4%), 1.6 ⁄ 2.1 mm (6.9% ⁄ 11.9%) and 1.0 ⁄ 2.0 mm (5.6% ⁄ 10.4%), with the new method smaller by 5.6–11.9%. 4. Conclusions and discussion
Fig. 7. Relationship between measurements made by our rendering image and the micrometer.
Based on a clinical ultrasound scanner and a spatial sensor, a new system for providing breast coronal images was successfully
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
developed in this paper. The tests on phantoms and three patients in vivo were conducted to preliminarily demonstrate the system feasibility. Unlike the existed 3D breast ultrasound scanner which requires a specially designed 3D probe [20–26], this method simply adds a spatial sensor on the routinely used 1D array linear probe. The clinician can scan the whole breast with the same probe to find all lesions. When coronal images are needed, the clinician only chooses images in lesion part to be reconstructed instead of switching probes. So this method can simplify the scanning procedure. The clinical result showed that the average scanning times for our system and the commercial 3D probe were 64 s and 74 s, respectively. This result demonstrated that this system could help to save examination time. In addition, this rendering method aimed at providing breast coronal images and it directly projected raw B-mode image data to the required coronal image. By omitting the volume reconstruction stage, this method could save rendering time and avoid one resampling process. In this system, to get rendering image, one important step was the definition of ROI manually. However, if this step can be finished automatically, this system will provide coronal images automatically and save examination time. Automatic ROI detection methods have been studied by researchers in different fields. In the compressed image data transmission, it was reported that automatic ROI detection was used to protect the important information [48]. To allow efficient compression for data storage, Ye et al. proposed an automatic method which identified ROI according to the edge information in high frequency sub-bands [49]. To segment dental X-ray images, Modi et al. presented an algorithm using region growing approach and Canny edge detector to automatically detect ROI [50]. It was also demonstrated that automated ROI selection was used in breast MRI images. This method firstly divided a lesion into several clusters and then calculated the most enhancement characteristics [51]. For the applications of automatic ROI in ultrasound images, Nguyen et al. reported a method using support vector machines (SVM) based top bound selection to visualize a fetal morphology on fetal volume image [52]. Compared with other types of image, B-images of breast is more difficult for automatic segmentation because of the irregular shape of breast tumors and image speckle noise. Su et al. reported an automatic ROI detection method for B-mode breast ultrasound images. In this method, after pre-processing, an image was divided into grids and features were detected. Then, according to the feature vectors, classification was performed to select ROI range [53]. Based on these methods, automatic detection will be developed for this new system and its performance on clinical data processing will be tested. In phantom experiments, a linear correlation (y = 1.0147x, R2 = 0.9927) was found between our system and the actual measurement values. The ICC values for intra- and inter-rater were 0.999 and 0.995, respectively. These results showed that the system had very good accuracy and repeatability. The average scanning time on three patients in vivo for our system and the commercial 3D probe were 64 s and 74 s, respectively. The results demonstrated that the new system required shorter scanning time in comparison with the procedure using the combination of the linear probe and the 3D probe. According to the preliminary results of the three lesions measured from the three subjects, the height and width of obtained by the new method was 5.6–11.9% smaller. One potential reason caused the differences was that the new method involved superimposing data of 2D images within a certain thickness to form the coronal image, leading to the loss of some details. While using the 3D probe, reslicing was used to obtain a coronal slice. However, it is difficult to confirm whether the image obtained by the 3D probe represented the true dimensions of the tumors. For the future clinical tests, the true tumor dimensions should be measured using another independent while accurate
433
method, such as MRI scanning or measurement based on tumors obtained in the surgery. This study focused on reporting the system development and the feasibility test. In spite of the limitations observed in the clinical experiment, the results of phantom experiments and clinical trials demonstrated that this new developed system was feasible for providing coronal images in clinical breast examinations, which was the aim of this study. A systematic largescale human test on this system has been planned. The improved clinical test method to overcome the above limitation will be used in this test. One challenge faced by any 3D ultrasound imaging for breast is the deformation of the breast during scanning by the probe. Using a 3D probe, it is put over the lesion and keeps stationary to acquire image using mechanical scanning inside the probe. The breast as well as the lesion may be compressed to some degree, but this compression is consistent during the whole image acquisition duration. While using free-hand 3D scanning, as used in the new system, the operator holds the linear probe to move over the lesion. The two scanning procedures may cause different deformations in breast and thus further induce variations in the lesion shape. Different approaches have been adopted to investigate the soft tissue deformation. An approach for solving the problem of soft tissue deformation is to use the non-rigid registration methods [54–56]. A sub-volume based method was developed to evaluate the ultrasound image deformations and validation experiments showed that this method could help to correct the deformations over 85% when this method was applied to register the distorted images with the original ones [57]. Xiao et al. used a block matching scheme and local statistics to register the breast deformations on ultrasound images. The registration error was presented less than 0.19 mm [58]. Khallaghi et al. introduced a feature-based registration algorithm on the basis of gradient information and image intensity. They compared their method with fast free-form, Demons, and B-spines registration methods and demonstrated its potential for recovering non-rigid deformations [59]. Researchers also demonstrated the combination of optical positional sensing device and non-rigid registration method to evaluate the deformations during the freehand 3D ultrasound scanning procedures [60,61]. The above results demonstrated that the non-rigid registration method could greatly correct deformation (over 85%) and had small registration error (less than 0.19 mm). Therefore, this method will be used in our system to overcome the influence of breast deformation. Acknowledgments This work was partially supported by the Hong Kong Innovative Technology Fund for HK-Guangdong Collaboration (GHP/047/09) and Hong Kong Polytechnic University (G-YL74). References [1] [2] [3] [4] [5] [6]
[7]
[8]
[9]
http://www.who.int/cancer/detection/breastcancer/en/index1.html. R. Siegel, D. Naishadham, A. Jemal. Cancer Statistics, 2012. pp. 11. http://www.breastcanceruk.org.uk/statistics/. American Cancer Society, Breast Cancer Facts & Figures, 2011–2012, American Cancer Society, Inc., Atlanta, 2011. pp. 17–19. W.E. Svensson, A review of the current status of breast ultrasound, Ultrasound Med. Biol. 6 (1997) 77–101. G.F. Xiao, 3-D free-hand ultrasound imaging and image analysis of the breast, PhD Dissertation. Department of Engineering Science, University of Oxford, 2001. S. Ciatto, M.R. Delturco, S. Catarzi, D. Morrone, The contribution of ultrasonography to the differential-diagnosis of breast-cancer, Neoplasma 41 (1994) 341–345. D. Rotten, J.M. Levaillant, L. Zerat, Analysis of normal breast tissue and of solid breast masses using three-dimensional ultrasound mammography, Ultrasound Obstet. Gynecol. 14 (1999) 114–124. L. Chen, Y. Chen, X.H. Diao, L. Fang, Y. Pang, A.Q. Cheng, W.P. Li, Y. Wang, Comparative study of automated breast 3-D ultrasound and handheld B-mode
434
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17] [18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26] [27] [28]
[29]
[30]
[31]
[32] [33]
[34]
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434 ultrasound for differentiation of benign and malignant breast masses, Ultrasound Med. Biol. 39 (2013) 1735–1742. Z.L. Wang, J.H. Xw, J.L. Li, Y. Huang, J. Tang, Comparison of automated breast volume scanning to hand-held ultrasound and mammography, Radiol. Med. 117 (2012) 1287–1293. D.O. Watermann, M. Foldi, A. Hanjalic-Beck, A. Hasenburg, A. Lughausen, H. Prompeler, G. Gitsch, E. Stickeler, Three-dimensional ultrasound for the assessment of breast lesions, Ultrasound Obstet. Gynecol. 25 (2005) 592–598. G.C. Meyberg-Solomayer, B. Kraemer, A. Bergmann, E. Kraemer, U. Krainick, D. Wallwiener, E.F. Solomayer, Does 3-D sonography bring any advantage to noninvasive breast diagnostics?, Ultrasound Med Biol. 30 (2004) 583–589. M. Tozaki, E. Fukuma, Accuracy of determining preoperative cancer extent measured by automated breast ultrasonography, Jpn. J. Radiol. 28 (2010) 771– 773. E. Baez, A. Huber, M. Vetter, B.J. Hackeloer, Minimal invasive complete excision of benign breast tumors using a three-dimensional ultrasound-guided mammotome vacuum device, Ultrasound Obstet. Gynecol. 21 (2003) 267–272. C. Lo, Y.W. Shen, C.S. Huang, R.F. Chang, Computer-aided multiview tumor detection for automated whole breast ultrasound, Ultrason. Imag. 36 (2014) 3– 17. C.J. Vyborny, T. Doi, K.F. O’Shaughnessy, H.M. Romsdahl, A.C. Schneider, A.A. Stein, Breast cancer: importance of spiculation in computer-aided detection, Radiology 215 (2000) 703–707. S.F. Huang, R.F. Chang, D.R. Chen, W.K. Moon, Characterization of spiculation on ultrasound lesions, IEEE Trans. Med. Imag. 23 (2004) 111–121. W.M. Chen, R.F. Chang, S.J. Kuo, C.S. Chang, W.K. Moon, S.T. Chen, D.R. Chen, 3D ultrasound texture classification using run difference matrix, Ultrasound in Medicine and Biology 31 (2005) 763–770. T. Tan, B. Platel, H. Huisman, C.I. Sanchez, R. Mus, N. Karssemeijer, ComputerAided Lesion Diagnosis in Automated 3-D Breast Ultrasound Using Coronal Spiculation, IEEE Trans. Med. Imag. 31 (2012) 1034–1042. D. Kotsianos-Hermle, S. Wirth, T. Fischer, K.M. Hiltawsky, M. Reiser, First clinical use of a standardized three-dimensional ultrasound for breast imaging, Eur. J. Radiol. 71 (2009) 102–108. D. Kotsianos-Hermle, K.M. Hiltawsky, S. Wirth, T. Fischer, K. Friese, M. Reiser, Analysis of 107 breast lesions with automated 3D ultrasound and comparison with mammography and manual ultrasound, Eur. J. Radiol. 71 (2009) 109–115. S.P. Sinha, M.A. Roubidoux, M.A. Helvie, A.V. Nees, M.M. Goodsitt, G.L. LeCarpentier, J.B. Fowlkes, C.L. Chalek, P.L. Carson, Multi-modality 3D breast imaging with X-ray tomosynthesis and automated ultrasound. in: Proceedings of the 29th Annual International Conference of the IEEE EMBS Cite Internationate, Lyon, France, August, 2007, pp. 23–26. X. Lin, J.W. Wang, F. Han, J.H. Fu, A.H. Li, Analysis of eighty-one cases with breast lesions using automated breast volume scanner and comparison with handheld ultrasound, Eur. J. Radiol. 81 (2012) 873–878. K.M. Kelly, J. Dean, W.S. Comulada, S.J. Lee, Breast cancer detection using automated whole breast ultrasound and mammography in radiographically dense breasts, Eur. Radiol. 20 (2010) 734–742. K.M. Kelly, J. Dean, S.J. Lee, W.S. Comulada, Breast cancer detection: radiologists’ performance using mammography with and without automated whole-breast ultrasound, Eur. Radiol. 20 (2010) 2557–2564. K.K. Shung, The principle of multidimensional arrays, Eur. J. Echocardiogr. 3 (2002) 149–153. T.R. Nelson, D.H. Pretorius, Three-dimensional ultrasound imaging, Ultrasound Med. Biol. 24 (1998) 1243–1270. M. Rao, T. Varghese, Correlation analysis of three-dimensional strain imaging using ultrasound two-dimensional array transducers, J. Acoust. Soc. Am. 124 (2008) 1858–1865. W.W. Jiang, A.H. Li, Y.P. Zheng, A semi-automated 3-D annotation method for breast ultrasound imaging: system development and feasibility study on phantoms, Ultrasound Med. Biol. 40 (2014) 434–446. C.D. Barry, C.P. Allott, N.W. John, P.M. Mellor, P.A. Arundel, D.S. Thomson, J.C. Waterton, Three-dimensional freehand ultrasound: image reconstruction and volume analysis, Ultrasound Med. Biol. 23 (1997) 1209–1224. Q.H. Huang, Y.P. Zheng, A new scanning approach for limb extremities using a water bag in freehand 3-D ultrasound, Ultrasound Med. Biol. 31 (2005) 575– 583. R.W. Prager, R.N. Rohling, A.H. Gee, L. Berman, Rapid calibration of 3-D freehand ultrasound, Ultrasound Med. Biol. 24 (1998) 855–869. A.K. Thittai, J.M. Yamal, J. Ophir, Small breast lesion classification performance using the normalized axial-shear strain area feature, Ultrasound Med. Biol. 39 (2013) 543–548. R. Prager, A. Gee, L. Berman, Stradx: real-time acquisition and visualization of freehand three-dimensional ultrasound, Med. Image Anal. 3 (1999) 129–140.
[35] G. Treece, R. Prager, A. Gee, L. Berman, Fast surface and volume estimation from non-parallel cross-sections, for freehand 3-D ultrasound, Med. Image Anal. 3 (1999) 141–173. [36] A. Gee, R. Prager, G. Treece, L. Berman, Narrow-band volume rendering for freehand 3D ultrasound, Comput Graph. 26 (2002) 463–476. [37] A. Fenster, D.B. Downey, 3-D ultrasound imaging: a review, IEEE Eng. Med. Biol. Mag. 15 (1996) 41–51. [38] O.V. Solberg, F. Lindseth, H. Torp, R.E. Blake, T.A.N. Hernes, Freehand 3D ultrasound reconstruction algorithms – a review, Ultrasound Med. Biol. 33 (2007) 991–1009. [39] Q.H. Huang, Y.P. Zheng, M.H. Lu, Z.R. Chi, Development of a portable 3D ultrasound imaging system for musculoskeletal tissues, Ultrasonics 43 (2005) 153–163. [40] Q.H. Huang, Y.P. Zheng, An adaptive squared-distance-weighted interpolation for volume reconstruction in 3D freehand ultrasound, Ultrasonics 44 (2006) e73–e77. [41] J.L. Gennisson, G. Cloutier, Sol-gel transition in agar-gelatin mixtures studied with transient elastography, IEEE Trans. Ultrason., Ferroelectr., Freq. Control 53 (2006) 716–723. [42] T.M. Mak, Y.P. Huang, Y.P. Zheng, Liver fibrosis assessment using transient elastography guided with real-time B-mode ultrasound imaging: a feasibility study, Ultrasound Med. Biol. 39 (2013) 956–966. [43] Y.P. Huang, Y.P. Zheng, S.F. Leung, A.F.T. Mak, Reliability of measurement of skin ultrasonic properties in vivo: a potential technique for assessing irradiated skin, Skin Res. Technol. 13 (2007) 55–61. [44] G. Rankin, M. Stokes, Reliability of assessment tools in rehabilitation: an illustration of appropriate statistical analyses, Clin. Rehab. 12 (1998) 187–199. [45] P.E. Shrout, J.L. Fleiss, Intraclass correlations: uses in assessing rater reliability, Psychol. Bull. 86 (1979) 420–428. [46] A.T. Stavros, Breast Ultrasound, Lippincott Williams & Wilkins, Philadelphia, 2004. pp. 46–47. [47] A.M. Dixon, Breast Ultrasound: How, Why and When, Churchill Livingstone, Edinburgh, 2008. pp. 94–99. [48] Z. Xue, K.K. Loo, J. Cosmas, M. Tun, L. Feng, P.Y. Yip, Error-resilient scheme for wavelet video codec using automatic ROI detection and Wyner–Ziv coding over packet erasure channel, IEEE Trans. Broadcast. 56 (2010) 481–493. [49] S.J. Ye, Y.F. Gu, Y. Zhan, An automatic ROI coding algorithm based on multiresolution target detection, in: 2008 2nd International Symposium on Systems and Control in Aerospace and Astronautics, vol. 1, issue: 2, 2008, pp. 610–613. [50] C.K. Modi, N.P. Desai, A simple and novel algorithm for automatic selection of ROI for dental radiograph segmentation. in: 24th Canadian Conference on Electrical and Computer Engineering, Niagara Falls, Canada, May 08–11, 2011. [51] M.J. Stoutjesdijk, J. Veltman, H. Huisman, N. Karssemeijer, J.O. Barentsz, J.G. Blickman, C. Boetes, Automated analysis of contrast enhancement in breast MRI lesions using mean shift clustering for ROI selection, J. Magn. Reson. Imag. 26 (2007) 606–614. [52] T.D. Nguyen, S.H. Kim, N.C. Kim, An automatic body ROI determination for 3D visualization of a fetal ultrasound volume, in: 9th International Conference on Knowledge-Based Intelligent Information and Engineering Systems, La Trobe University, Melbourne, Australia, September 14–16, 2005. [53] Y. Su, Y. Wang, J. Jiao, Y. Guo, Automatic detection and classification of breast tumors in ultrasonic images using texture and morphological features, Open Med. Inform. J. S1–3 (2011) 26–37. [54] W.R. Crum, T. Hartkens, D.L.G. Hill, Non-rigid image registration: theory and practice, Br. J. Radiol. 77 (2004) S140–S153. [55] Y.J. Guo, R. Sivaramakrishna, C.C. Lu, J.S. Suri, S. Laxminarayan, Breast image registration techniques: a survey, Med. Biol. Eng. Comput. 44 (2006) 15–26. [56] C. Leung, K. Hashtrudi-Zaad, P. Foroughi, P. Abolmaesumi, A real-time intrasubject elastic registration algorithm for dynamic 2-D ultrasound images, Ultrasound Med. Biol. 35 (2009) 1159–1176. [57] J.F. Krucker, G.L. LeCarpentier, J.B. Fowlkes, P.L. Carsen, Rapid elastic image registration for 3-D ultrasound, IEEE Trans. Med. Imag. 21 (2002) 1384–1394. [58] G.F. Xiao, J.M. Brady, J.A. Noble, M. Burcher, R. English, Nonrigid registration of 3-D free-hand ultrasound images of the breast, IEEE Trans. Med. Imag. 21 (2002) 405–412. [59] S. Khallaghi, C.G.M. Leung, K. Hastrudi-Zaad, P. Foroughi, C. Nguan, P. Abolmaesumi, Experimental validation of an intrasubject elastic registration algorithm for dynamic-3D ultrasound images, Med. Phys. 39 (2012) 5488– 5497. [60] A.H. Gee, G.M. Treece, R.W. Prager, C.J.C. Cash, L. Berman, Rapid registration for wide field of view freehand three-dimensional ultrasound, IEEE Trans. Med. Imag. 22 (2003) 1344–1357. [61] G.M. Treece, R.W. Prager, A.H. Gee, L. Berman, Correction of probe pressure artifacts in freehand 3D ultrasound, Med. Image Anal. 6 (2002) 199–214.
Contents lists available at ScienceDirect
Ultrasonics journal homepage: www.elsevier.com/locate/ultras
A novel breast ultrasound system for providing coronal images: System development and feasibility study Wei-wei Jiang a, Cheng Li b,c, An-hua Li b, Yong-Ping Zheng a,⇑ a
Interdisciplinary Division of Biomedical Engineering, The Hong Kong Polytechnic University, Kowloon, Hong Kong, China Department of Ultrasound, State Key Laboratory of Oncology in Southern China, Sun Yat-Sen University Cancer Center, Guangzhou, China c Department of Ultrasound, Hospital of Traditional Chinese Medicine of Zhongshan, Zhongshan, China b
a r t i c l e
i n f o
Article history: Received 9 May 2014 Received in revised form 10 September 2014 Accepted 16 September 2014 Available online 27 September 2014 Keywords: Coronal image Breast ultrasound Image rendering Breast cancer Breast diagnosis
a b s t r a c t Breast ultrasound images along coronal plane contain important diagnosis information. However, conventional clinical 2D ultrasound cannot provide such images. In order to solve this problem, we developed a novel ultrasound system aimed at providing breast coronal images. In this system, a spatial sensor was fixed on an ultrasound probe to obtain the image spatial data. A narrow-band rendering method was used to form coronal images based on B-mode images and their corresponding spatial data. Software was developed for data acquisition, processing, rendering and visualization. In phantom experiments, 20 inclusions with different size (5–20 mm) were measured using this new system. The results obtained by the new method well correlated with those measured by a micrometer (y = 1.0147x, R2 = 0.9927). The phantom tests also showed that this system had excellent intra- and inter-operator repeatability (ICC > 0.995). Three subjects with breast lesions were scanned in vivo using this new system and a commercially available three-dimensional (3D) probe. The average scanning times for the two systems were 64 s and 74 s, respectively. The results revealed that this new method required shorter scanning time. The tumor sizes measured on the coronal plane provided by the new method were smaller by 5.6–11.9% in comparison with the results of the 3D probe. The phantom tests and preliminary subject tests indicated the feasibility of this system for clinical applications by providing additional information for clinical breast ultrasound diagnosis. Ó 2014 Elsevier B.V. All rights reserved.
1. Introduction Breast cancer is the most common cancer in women worldwide. In the Global Health Estimates of World Health Organization (WHO), it was estimated that 508,482 women died of breast cancer in 2011 in the world [1]. In America, it was reported that 226,870 women were diagnosed with breast cancer and 39,510 of them died of breast cancer in 2012 [2]. According to the report of Breast Cancer UK, breast cancer accounted for 31% of cancers diagnosed in women [3]. Up to now, there has not been an effective method to prevent breast cancer and early detection has remained the cornerstone for breast cancer control [4]. Among all breast cancer detection methods, ultrasound plays an important role in breast cancer deaths decline for its advantages of radiation-free, real-time and ⇑ Corresponding author at: Interdisciplinary Division of Biomedical Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, China. Tel.: +852 27667664; fax: +852 23624365. E-mail address: [email protected] (Y.-P. Zheng). http://dx.doi.org/10.1016/j.ultras.2014.09.009 0041-624X/Ó 2014 Elsevier B.V. All rights reserved.
suitable for dense breast [5,6]. Ultrasound has long been recognized as a valuable tool to distinguish between cysts and solid masses. With the rapid development of ultrasound techniques and greatly increased images quality, breast ultrasound can now not only be used for characterizing cysts, but also differentiating benign from malignant lesions. In a breast abnormalities (259 carcinomas, 1820 benign) examination, ultrasound could help to avoid unnecessary biopsy with benign diagnosis results in 71 suspicious cases at palpation or mammography [7]. Therefore, routine ultrasound examination can help to reduce unnecessary biopsies. In clinical breast ultrasound examination, 2D ultrasound probe is routinely used which can only provide transverse and longitudinal images but no coronal images. However, information on this plane has been proved to be beneficial for clinical diagnosis [8– 14]. Rotten et al. analyzed images of normal breast tissue and breast lesions and found four diagnosis features on coronal plane [8]. Among these features, one was defined as compressive pattern which was thought to be associated with benign lesions. In this pattern, the continuous hyperechoic bands of tissue peripheral to
428
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
the masses appeared to be distinct from the central part. Another feature was called converging pattern which was a typical characteristic of malignant lesions. For this pattern, a stellate distortion consisting of alternating hypoechoic and hyperechoic lines converged towards to the hypoechoic central masses [8]. In the study of Chen et al., the two features were described as hyperechoic rim and retraction phenomenon. They were used independently to differentiate breast lesions and good accuracy (95.9% for hyperechoic rim; 96.8% for retraction phenomenon) and specificity (92.8% for hyperechoic rim; 100% for retraction phenomenon) were reported [9]. In fact, the feature of retraction phenomenon on coronal plane, also called spiculation, was reported by other researchers with high specificity (98.4% in [10]; 94.6% in [11]) in tumor diagnosis. In the report of Meyberg-Solomayer et al., the coronal plane could provide tumor classification information when the infiltrative zone was not visible (17 of 39 cases) or unclear (6 of 8 cases) in 2D ultrasound imaging [12]. This result demonstrated that the image on this plane could offer a better assessment when the infiltrative zone surrounding the lesion was unclear or not visible on conventional 2D images, which could help to reduce biopsies. Images along coronal plane were also beneficial for tumor extent measurement [13] and ultrasound-guided vacuum-assisted core-needle biopsy [14]. Based on images on coronal images, various computer-aided diagnosis (CAD) methods were presented to help to automatically detect tumor candidate [15], mark spiculated masses [16,17], and classify tumor stages [18,19]. Breast coronal images can be provided by the technique of three-dimensional (3D) ultrasound imaging. Many researchers have been studying on this technique. One approach of 3D breast ultrasound imaging was to scan the breast using the conventional 2D probe, which was driven by a mechanical motor [20–22]. A typical representation of this approach was the method proposed by Kotsianos-Hermle et al. [20]. In this method, breast was compressed by two paddles and a probe was driven mechanically on the top of the paddle to acquire the breast images. Another approach was to scan the patient with a specially designed probe when the patient was in supine position, such as the commercial products Automated Breast Volume Scanner (ABVS) of Siemens and the Automated Whole-Breast Ultrasound (AWBU) of Sonocine [23–25]. Another approach for 3D ultrasound imaging is the 2D array ultrasound probe which uses electronic pyramidal scanning [26]. They have been used successfully for real-time 3D imaging of the heart, where high volume frame rate (40 volumes per second) is required [27], but are seldom used for breast imaging. A high volume frame is not as necessary for breast imaging, unless breast tissue motion needs to be tracked in three dimensions in real time, as in 3D elastography [28]. Therefore, the first two approaches of 3D ultrasound imaging can provide breast coronal images. However, in these methods, the driven motor or specially designed probes were required for scanning. These equipments were bulky and large, which were inconvenient for clinical scanning to offer regular motions. In addition, the moving manner of the probe in these systems was predefined so the operator could not move the probe to the desired position freely. Some regions such as axillary region and tissue against the chest wall were not accessible by using these systems. Therefore, there are still many works to be done before 3D ultrasound imaging technique can be widely used in clinical breast examination. 2D ultrasound imaging remains the dominant scanning mode for clinical breast ultrasound diagnosis. Accordingly, this study was aimed to develop a breast ultrasound system for providing coronal images based on the clinical 2D ultrasound scanner. In the following sections, this system is described in details. The system accuracy and reliability tests based on phantoms are presented. Preliminary clinical tests were also performed to demonstrate the system feasibility.
2. Methods 2.1. System overview A corresponding freehand 3D ultrasound annotation system was previously developed and successfully used for annotating breast ultrasound images [29]. Fig. 1 shows the diagram of this system. It consisted of three main components: an ultrasound machine (EUB-8500, Hitachi, Tokyo, Japan) with a linear 2D probe (EUP-L65/6–14 MHz, Hitachi, focused probe, 6–14 MHz), an electromagnetic spatial sensing device (med-SAFE, Ascension Technology, Burlington, VT, USA) and a computer with Intel Core i5 3.35 GHz CPU and 3.5 GB of memories. A video capture card (NIIMAQ PCI/PXI-1411, National Instruments Corporation, Austin, TX, USA) and a customized program were installed on this computer. The electromagnetic spatial sensing device was employed to acquire the image spatial data in this system. The selected device had high spatial accuracy. The documented positional accuracy and angular accuracy of this device were 1.4 mm and 0.5°, respectively (medSAFE Manual, Ascension Technology). The spatial device was comprised of a control box, transmitter and a sensor. The diameter of the cylindric sensor was 2.0 mm and the length was 9.9 mm. This sensor dimension was small so it was easy to be fixed on the ultrasound probe by a custom-designed kit. The image spatial data acquired by this sensor included three positions (x, y, z) and three orientations (azimuth, elevation, roll). These data were sent from the control box of the spatial device to the computer through its serial port. The sampling rate of medSAFE was 100 Hz, which was higher than the ultrasound imaging rate. So sufficient spatial data were collected and averaging was used to improve the accuracy of the system on distance and angle measurement. During scanning, the video stream of 2D B-mode ultrasound images was captured by the video capture card and sent to the computer. Meanwhile, the spatial data of these images were also sent to the computer by the control box of the spatial device. The developed program acquired and recorded these images together with their corresponding spatial information for the further visualization and rendering. Spatial calibration experiments were performed for the system to determine the position and orientation offsets between the ultrasound image and the spatial sensor. A cross-wire phantom was used to calibrate this system [30,31]. Two wires were crossed and submerged in a water tank and the wire ends were fixed to the tank. The ultrasound probe was moved slowly to scan the wire cross. If an image with a clear cross was found, this image and its spatial information would be recorded. In each experiment, 60 images from various directions were captured. According to the
Fig. 1. Diagram of the breast ultrasound rendering system developed in this study, which consisted of an ultrasound machine, a spatial sensing device and a computer.
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
pixel position of the cross on each image and the positional data read from the spatial sensor, the spatial transformation was then calculated using the Levenberg–Marquardt nonlinear algorithm [32]. Levenberg–Marquardt nonlinear algorithm was a common technique to solve nonlinear least squares problems. This technique took advantages of both the gradient descent and Gauss– Newton methods. It acted more like a gradient-descent method when the parameters were far from their optimal values, and acted more like the Gauss–Newton method when the parameters were close to their optimal values. This method was robust and popular in solving nonlinear least squares problems so it was used in this study. 2.2. Data acquisition and volume rendering A program for this system was developed using Visual C++ (Microsoft, Redmond, WA, USA) and Visualization Toolkits (VTk, Kitware Inc., NY, USA). The software interface is shown in Fig. 2, which consisted of control bar and the display window. The main functions on the control bar could be organized into four parts: data acquisition, data processing, image rendering and visualization. The data acquisition part was designed for the capture controlling of ultrasound images and spatial data. The acquisition was in real-time mode. Acquisition parameters such as image size could be set by the operator in this part. The processing part contained functions to manually choose valid image frames and to select region of interest (ROI) after data were captured. The processing functions were conducted in off-line mode. The rendering function was responsible for the volume rendering to obtain the coronal images. For this system, the off-line rendering procedure normally took less than 10 s. After rendering, the coronal image was displayed on the display window. This procedure was finished by the visualization function, which took less than 1 s. The visual-
429
ization function also included the real-time display of the original ultrasound images during the acquisition. During the scanning, the operator held the probe and moved freely on the subject. Ultrasound images together with their corresponding spatial data were displayed on the software interface in real time, also shown in Fig. 2. To control the image acquisition during scanning the subject by hands, a foot switch was designed. When the operator stepped on the switch, images with their corresponding spatial data were recorded. If the operator stepped again, the recording was stopped. During scanning, if features of breast tumor such as irregular border were found on images, the operator could step on the foot switch to record images and spatial data for the further processing. For different operators, the scanning speeds were different. This had influence to the reconstructed image resolution along the scanning direction as the frame rate was fixed, which was 21 frames per second for our system. When the probe was moved faster, the distance between two consecutive 2D images would be larger, thus lower resolution along the scanning direction. According to the frame rate collecting B-mode images, the probe could be moved as fast as 10 mm/s to achieve a gap smaller than 0.5 mm. Thus, using a normal scanning speed, the scanning gaps would be very small and be filled through the later interpolation. Furthermore, after scanning, the acquired images could be displayed in 3D space so that the operator could investigate if there was region not covered. If obvious gaps were found, this data sequence would be discarded and the operator may scan the lesion region again. In clinical breast ultrasound images, since the useful information such as lesion region was only small part on raw 2D ultrasound images, especially for small lesions (<10 mm diameter) [33]. In clinical breast imaging, small lesions were more easily missed by the clinicians. Therefore, in this study, the region of interest (ROI) should be defined before rendering. In the present
Fig. 2. The software interface of the breast ultrasound rendering system comprising the control panel and display window.
430
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
Fig. 3. The method used for defining region of interest for rendering.
method, the start and end B-mode image including diagnosis information were firstly marked by the operator when reviewing images, as shown in Fig. 3. To get the region of interest from original images, the operator could set two points on one image to define a region including the lesion. On the basis of the two points, the program could automatically calculate the region of interest and discard the information outside this region, as shown in Fig. 3. The conventional volume rendering methods usually contained two steps. The first step was to reconstruct a regular voxel array using the recorded raw images. Then the voxel array was visualized using the planar reslicing. The two steps involve two resampling stages, from raw image pixels to the voxel array, and from the voxel array to the slice pixels. However, for the coronal imaging purpose of this study, the first step to form the voxel array was not necessary. In this study, a volume rendering approach, which obtained the reslice plane with a certain thickness directly from the raw data, was used. This approach only required one resampling stage. Since the resampling procedure inevitably involves the data approximation, this approach was expected to provide more accurate visualization. This rendering method, also called narrow-band volume rendering, was first introduced by Gee et al. [34–36]. Since the coronal image was endowed with a certain thickness based on the raw image data, the narrow-band meant a layer with limited thickness and the value of thickness could be determined by the operator. In this study, the thickness ranged from 4.8 mm to 9.7 mm. In this volume rendering method, as shown in Fig. 4, all selected 2D images were projected to the coronal plane to form a new 2D coronal image. First, according to the
Fig. 4. An illustration of the volume rendering method in this developed system.
scanning region and the size of ROI, the coronal image coordinate system with a regular pixel array was defined. The density of the matrix was determined by operators during reconstruction, while the resolution of 2D imaging plane was determined by the ultrasound transducer. On the basis of the spatial data and the calibration matrix, each pixel on the original 2D images was transformed to the new coordinate system. The final value of the pixel intensity on the new coronal image was calculated based on all pixels falling into its region. Two methods including minimum intensity compounding and averaging were used to calculate the intensity. The minimum intensity compounding was to choose the minimum intensity from all pixels falling into the pixel region as the final value of this pixel. The average method was to average all pixels in the region and the average value was used as the final value. Minimum intensity compounding was suitable for the fluid-filled lesion such as cyst. Simple averaging could produce an X-ray like image [36]. After the data mapping, the intensity of most pixels could be calculated. However, in this method, the free-hand scanning was adopted. This scanning method might cause gaps though it had special advantages such as the flexibility of manipulation [27,37]. In order to fill gaps, interpolation was used to compute all empty grids. A variety of methods, such as the nearest neighbor interpolation and linear interpolation, had been reported for the gap filling [38–40]. In this system, the bilinear interpolation was used to produce the trade-off between the interpolation performance and computation time. For each empty pixel, its intensity was the average of non-empty pixels in the nearest 2-by-2 neighborhood. 2.3. Validation phantom and experiments To demonstrate the system accuracy and repeatability in providing coronal images, phantoms were designed for validation experiments. Today, various commercial tissue phantoms are available for different medical applications. However, in this study, phantoms containing 20 sphere inclusions with different dimensions were needed. There are no commercial phantoms available with all these different sizes of inclusions. Therefore, the customdesigned phantoms were employed in this study. The diagram of the phantom was shown in Fig. 5(a), which comprised of matrix and inclusion. In order to mimic the clinical ultrasound images of breast tumor, the material of matrix should be hyperecho and the inclusion should be hypoecho. The inclusion was made from a material called crystal soil which was a water-absorbent polymer. The crystal soil could hold up to 180 times it volume in water therefore the main component of the inclusion was water, as shown in Fig. 5(b). Fig. 5(c) is the matrix which was a mixture of water, 5% gelatin (Gelatin from porcine skin, Type A, Sigma Aldrich, St. Louis, MO, USA) and 3% agar (Agar, granular powder, general purpose grade, Fisher Scientific, Loughborough, UK). The agar was used as the scatter to make the matrix in hyperecho status. The mixture was initially heated to 50 °C. This temperature could keep the mixture in the liquid status and the agar was not dissolved so the agar particles could provide speckles. And then the temperature was reduced to 24 °C quickly to avoid agar sedimentation because the mixture began to solidify at 24 °C [41,42]. Before the matrix was totally solidified, the inclusion was put into the matrix using tweezers. And then the mixture with inclusion was put into the refrigerator for 12 h for the further solidification. The ultrasound image of the phantom is shown in Fig. 5(d), which was similar to the clinical images of breast cancer, especially the cyst cases. In validation experiments, the accuracy and repeatability of this system were tested. For accuracy test, 20 inclusions with different size (5–20 mm) were used. The diameters of these inclusions were firstly measured by a micrometer. For each inclusion, the measure-
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
431
Fig. 5. The validation phantom. (a) The diagram of validation phantom; (b) the inclusion of crystal soil; (c) the phantom matrix; (d) the ultrasound image of the phantom.
ment was repeated three times and the average was used as the true size of the inclusion. Then these inclusions were put into matrix. Totally five phantoms were made and each of them included four inclusions. These phantoms were scanned by the system and coronal images were obtained after volume rendering. The inclusion dimension was then measured on the coronal images, here the coronal plane refers to the plane parallel to the phantom surface, which are similar to the coronal images in breast ultrasound imaging. The measurement results were compared with the real size of these inclusions to test the system accuracy. For reliability tests, the operator held the probe to scan the phantom and then coronal images were reconstructed to measure the inclusions size. In intra-rater tests, one operator scanned these phantoms twice at the intervals of at least 5 h. For inter-rater tests, the other operator performed the same procedure on phantoms after the first operator finished. The parameter of Intraclass Correlation Coefficient (ICC) was used to analyze data [43,44]. The intrarater and inter-rater repeatability was, respectively, assessed by the ICC (2, 1) and ICC (3, 1) [45]. All the statistical analyses were conducted by the statistical software SPSS (SPSS Inc., Chicago, USA). 2.4. In vivo experiments Although phantom experiments could provide evidence for the feasibility of this new method, in vivo experiments were more important to demonstrate the system performance on the clinical application. In the experiments, a commercial ultrasound machine (Logiq E9, GE Healthcare, New York, USA) with a linear 2D probe (ML6-15-D, GE Healthcare, New York, USA) and a 3D probe (RSP6-16-D, GE Healthcare, New York, USA) was employed to test the feasibility of the new system. Since the ultrasound scanner used in our system did not have a 3D probe, the above commercial scanner which had a 3D probe was used in the in vivo experiments. The aim of the subject tests was to compare the performance of our system with that of a commercial 3D probe. This 3D probe was a routinely used in clinics so it met our requirement. Three patients (women, 63, 57, and 53 years old), who had breast pain or findings using palpation, were recruited to have ultrasonography. After ultrasound examinations, the three patients had biopsies. The histopathological results showed that they were all ductal carcinoma.
During the scanning by our system, the patient lay on the bed in supine position. An operator held the probe to scan the breast in the orthogonal antiradial scanning pattern [46,47]. When a lesion was found, the operator would not stop scanning but remember the lesion location until covering two whole breasts. This step could help to ensure that all breast regions were covered. After two breasts were scanned, the operator moved the probe to the lesion region to scan this part slowly and saved images together with their corresponding annotation information. The above scanning procedure was the routine clinical breast ultrasound examination procedure. The total scanning time, including the time spending on scanning two breasts and scanning the lesion region, was recorded. After the scanning using the new system, the patient kept the same posture on the bed and was scanned using a combination of the linear probe and the GE mechanically scanning 3D probe. The operator held the linear probe and scanned the two breasts to find the lesion region. Then she changed to the 3D probe and put the probe on the lesion part. The 3D probe was kept stationary on the lesion for about 15 s until the acquisition was finished. The scanning time for this procedure was also recorded. The in vivo experiments were conducted in the Sun Yat-Sen University Cancer Center (Guangzhou, China). The patients were given the informed consent before scanning, which was approved by the human subject ethics committee of the institution. To demonstrate the performance of the new system in clinics, the required scanning time and the formed coronal images of the two methods were compared. In this study, only the scanning time, which was spent on acquiring images, was compared. The time spending on the coronal image reconstruction was not included because it was in the review section in clinical breast examination. The review section was off-line. Therefore, compared with the time spending on review section, the scanning time was more important. The time saved during scanning section could help to scan more patients in clinical examinations. For the commercially available GE 3D probe, after image reconstruction, slices along the coronal plane could be provided for different depths. After reviewing all slices, the plane with largest dimension in tumor size was selected and the width and height of the tumor were measured. The depth of the selected slice was also recorded. Using the new system, the coronal image at the same depth was reconstructed. The tumor
432
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
size was also measured on this image and compared with the GE 3D scanner results.
Table 1 Comparison of scanning time of the two methods.
3. Results 3.1. Validation results
By 3D probe By new system
Patient 1 (s)
Patient 2 (s)
Patient 3 (s)
Average time (s)
75 62
73 66
75 63
74.3 63.7
Fig. 6 is a typical coronal image of the phantom. The dark part indicated the inclusion. The dimensional measurement was also shown on the image. The measurement results were compared with the actual sizes of inclusions to demonstrate the system accuracy. Fig. 7 illustrates the correlation about the phantom diameters measured by the 3D ultrasound rendering image and the micrometer. The results demonstrated a very good linear correlation (y = 1.0147x, R2 = 0.9927) between the results obtained by the two methods. This indicated that the measurement using the new system well matched the actual values. The ICC values for intra- and inter-rater were 0.999 and 0.995, respectively. The high values indicate excellent repeatability of the new system. 3.2. Results for the in vivo experiments The scanning time for the two methods is summarized in Table 1. Using the commercial linear probe together with the 3D probe, the average scanning time was 74 s. And for the new system, the mean acquisition time was 64 s. Compared with the commercial scanner, this new system required shorter scanning time.
Fig. 8. The comparison of coronal images from two methods. (a) Coronal image from GE 3D scanner; (b) coronal image obtained by our rendering system.
Table 2 The tumor size comparison between the two methods.
By 3D probe By new system Difference
Fig. 6. The typical coronal image and dimensional measurement of phantom for accuracy test, here the coronal plane refers to a plane in parallel to the phantom surface.
Patient 1 (mm)
Patient 2 (mm)
Patient 3 (mm)
Width
Height
Width
Height
Width
Height
18.9
19.3
23.3
17.7
18.0
19.2
17.7
17.3
21.7
15.6
17.0
17.2
1.2 (5.7%)
2.0 (10.4%)
1.6 (6.9%)
2.1 (11.9%)
1.0 (5.6%)
2.0 (10.4%)
The reason for the longer time of the commercial scanner was probably due to the switch between probes. In clinical breast examination, two steps are usually needed. The first step is scanning the whole breast to find all lesions. For this step, the linear probe ML6-15-D was used and it could allow the operator to investigate tumor in real-time on the screen. Then the operator put down the linear probe and switched to the 3D probe to image the tumor region. The probe switching obviously required additional scanning time and caused inconvenience for the examination. Fig. 8(a) shows the coronal image obtained using the GE 3D probe. As shown in the figure, the clinician manually put cross on the image to measure the tumor width and height. After setting depth and thickness, the coronal image on the same position could be provided by the rendering method of the new system, as shown in Fig. 8(b). The tumor size could also be measured by the operator on this image. The comparison of tumor size between the two methods for three patients is summarized in Table 2. As shown in Fig. 8, the tumor size is represented by width ⁄ height in mm. For the three lesions, the differences between the results obtained by the new method and the GE scanner were 1.2 ⁄ 2.0 mm (5.7% ⁄ 10.4%), 1.6 ⁄ 2.1 mm (6.9% ⁄ 11.9%) and 1.0 ⁄ 2.0 mm (5.6% ⁄ 10.4%), with the new method smaller by 5.6–11.9%. 4. Conclusions and discussion
Fig. 7. Relationship between measurements made by our rendering image and the micrometer.
Based on a clinical ultrasound scanner and a spatial sensor, a new system for providing breast coronal images was successfully
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434
developed in this paper. The tests on phantoms and three patients in vivo were conducted to preliminarily demonstrate the system feasibility. Unlike the existed 3D breast ultrasound scanner which requires a specially designed 3D probe [20–26], this method simply adds a spatial sensor on the routinely used 1D array linear probe. The clinician can scan the whole breast with the same probe to find all lesions. When coronal images are needed, the clinician only chooses images in lesion part to be reconstructed instead of switching probes. So this method can simplify the scanning procedure. The clinical result showed that the average scanning times for our system and the commercial 3D probe were 64 s and 74 s, respectively. This result demonstrated that this system could help to save examination time. In addition, this rendering method aimed at providing breast coronal images and it directly projected raw B-mode image data to the required coronal image. By omitting the volume reconstruction stage, this method could save rendering time and avoid one resampling process. In this system, to get rendering image, one important step was the definition of ROI manually. However, if this step can be finished automatically, this system will provide coronal images automatically and save examination time. Automatic ROI detection methods have been studied by researchers in different fields. In the compressed image data transmission, it was reported that automatic ROI detection was used to protect the important information [48]. To allow efficient compression for data storage, Ye et al. proposed an automatic method which identified ROI according to the edge information in high frequency sub-bands [49]. To segment dental X-ray images, Modi et al. presented an algorithm using region growing approach and Canny edge detector to automatically detect ROI [50]. It was also demonstrated that automated ROI selection was used in breast MRI images. This method firstly divided a lesion into several clusters and then calculated the most enhancement characteristics [51]. For the applications of automatic ROI in ultrasound images, Nguyen et al. reported a method using support vector machines (SVM) based top bound selection to visualize a fetal morphology on fetal volume image [52]. Compared with other types of image, B-images of breast is more difficult for automatic segmentation because of the irregular shape of breast tumors and image speckle noise. Su et al. reported an automatic ROI detection method for B-mode breast ultrasound images. In this method, after pre-processing, an image was divided into grids and features were detected. Then, according to the feature vectors, classification was performed to select ROI range [53]. Based on these methods, automatic detection will be developed for this new system and its performance on clinical data processing will be tested. In phantom experiments, a linear correlation (y = 1.0147x, R2 = 0.9927) was found between our system and the actual measurement values. The ICC values for intra- and inter-rater were 0.999 and 0.995, respectively. These results showed that the system had very good accuracy and repeatability. The average scanning time on three patients in vivo for our system and the commercial 3D probe were 64 s and 74 s, respectively. The results demonstrated that the new system required shorter scanning time in comparison with the procedure using the combination of the linear probe and the 3D probe. According to the preliminary results of the three lesions measured from the three subjects, the height and width of obtained by the new method was 5.6–11.9% smaller. One potential reason caused the differences was that the new method involved superimposing data of 2D images within a certain thickness to form the coronal image, leading to the loss of some details. While using the 3D probe, reslicing was used to obtain a coronal slice. However, it is difficult to confirm whether the image obtained by the 3D probe represented the true dimensions of the tumors. For the future clinical tests, the true tumor dimensions should be measured using another independent while accurate
433
method, such as MRI scanning or measurement based on tumors obtained in the surgery. This study focused on reporting the system development and the feasibility test. In spite of the limitations observed in the clinical experiment, the results of phantom experiments and clinical trials demonstrated that this new developed system was feasible for providing coronal images in clinical breast examinations, which was the aim of this study. A systematic largescale human test on this system has been planned. The improved clinical test method to overcome the above limitation will be used in this test. One challenge faced by any 3D ultrasound imaging for breast is the deformation of the breast during scanning by the probe. Using a 3D probe, it is put over the lesion and keeps stationary to acquire image using mechanical scanning inside the probe. The breast as well as the lesion may be compressed to some degree, but this compression is consistent during the whole image acquisition duration. While using free-hand 3D scanning, as used in the new system, the operator holds the linear probe to move over the lesion. The two scanning procedures may cause different deformations in breast and thus further induce variations in the lesion shape. Different approaches have been adopted to investigate the soft tissue deformation. An approach for solving the problem of soft tissue deformation is to use the non-rigid registration methods [54–56]. A sub-volume based method was developed to evaluate the ultrasound image deformations and validation experiments showed that this method could help to correct the deformations over 85% when this method was applied to register the distorted images with the original ones [57]. Xiao et al. used a block matching scheme and local statistics to register the breast deformations on ultrasound images. The registration error was presented less than 0.19 mm [58]. Khallaghi et al. introduced a feature-based registration algorithm on the basis of gradient information and image intensity. They compared their method with fast free-form, Demons, and B-spines registration methods and demonstrated its potential for recovering non-rigid deformations [59]. Researchers also demonstrated the combination of optical positional sensing device and non-rigid registration method to evaluate the deformations during the freehand 3D ultrasound scanning procedures [60,61]. The above results demonstrated that the non-rigid registration method could greatly correct deformation (over 85%) and had small registration error (less than 0.19 mm). Therefore, this method will be used in our system to overcome the influence of breast deformation. Acknowledgments This work was partially supported by the Hong Kong Innovative Technology Fund for HK-Guangdong Collaboration (GHP/047/09) and Hong Kong Polytechnic University (G-YL74). References [1] [2] [3] [4] [5] [6]
[7]
[8]
[9]
http://www.who.int/cancer/detection/breastcancer/en/index1.html. R. Siegel, D. Naishadham, A. Jemal. Cancer Statistics, 2012. pp. 11. http://www.breastcanceruk.org.uk/statistics/. American Cancer Society, Breast Cancer Facts & Figures, 2011–2012, American Cancer Society, Inc., Atlanta, 2011. pp. 17–19. W.E. Svensson, A review of the current status of breast ultrasound, Ultrasound Med. Biol. 6 (1997) 77–101. G.F. Xiao, 3-D free-hand ultrasound imaging and image analysis of the breast, PhD Dissertation. Department of Engineering Science, University of Oxford, 2001. S. Ciatto, M.R. Delturco, S. Catarzi, D. Morrone, The contribution of ultrasonography to the differential-diagnosis of breast-cancer, Neoplasma 41 (1994) 341–345. D. Rotten, J.M. Levaillant, L. Zerat, Analysis of normal breast tissue and of solid breast masses using three-dimensional ultrasound mammography, Ultrasound Obstet. Gynecol. 14 (1999) 114–124. L. Chen, Y. Chen, X.H. Diao, L. Fang, Y. Pang, A.Q. Cheng, W.P. Li, Y. Wang, Comparative study of automated breast 3-D ultrasound and handheld B-mode
434
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17] [18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26] [27] [28]
[29]
[30]
[31]
[32] [33]
[34]
W.-w. Jiang et al. / Ultrasonics 56 (2015) 427–434 ultrasound for differentiation of benign and malignant breast masses, Ultrasound Med. Biol. 39 (2013) 1735–1742. Z.L. Wang, J.H. Xw, J.L. Li, Y. Huang, J. Tang, Comparison of automated breast volume scanning to hand-held ultrasound and mammography, Radiol. Med. 117 (2012) 1287–1293. D.O. Watermann, M. Foldi, A. Hanjalic-Beck, A. Hasenburg, A. Lughausen, H. Prompeler, G. Gitsch, E. Stickeler, Three-dimensional ultrasound for the assessment of breast lesions, Ultrasound Obstet. Gynecol. 25 (2005) 592–598. G.C. Meyberg-Solomayer, B. Kraemer, A. Bergmann, E. Kraemer, U. Krainick, D. Wallwiener, E.F. Solomayer, Does 3-D sonography bring any advantage to noninvasive breast diagnostics?, Ultrasound Med Biol. 30 (2004) 583–589. M. Tozaki, E. Fukuma, Accuracy of determining preoperative cancer extent measured by automated breast ultrasonography, Jpn. J. Radiol. 28 (2010) 771– 773. E. Baez, A. Huber, M. Vetter, B.J. Hackeloer, Minimal invasive complete excision of benign breast tumors using a three-dimensional ultrasound-guided mammotome vacuum device, Ultrasound Obstet. Gynecol. 21 (2003) 267–272. C. Lo, Y.W. Shen, C.S. Huang, R.F. Chang, Computer-aided multiview tumor detection for automated whole breast ultrasound, Ultrason. Imag. 36 (2014) 3– 17. C.J. Vyborny, T. Doi, K.F. O’Shaughnessy, H.M. Romsdahl, A.C. Schneider, A.A. Stein, Breast cancer: importance of spiculation in computer-aided detection, Radiology 215 (2000) 703–707. S.F. Huang, R.F. Chang, D.R. Chen, W.K. Moon, Characterization of spiculation on ultrasound lesions, IEEE Trans. Med. Imag. 23 (2004) 111–121. W.M. Chen, R.F. Chang, S.J. Kuo, C.S. Chang, W.K. Moon, S.T. Chen, D.R. Chen, 3D ultrasound texture classification using run difference matrix, Ultrasound in Medicine and Biology 31 (2005) 763–770. T. Tan, B. Platel, H. Huisman, C.I. Sanchez, R. Mus, N. Karssemeijer, ComputerAided Lesion Diagnosis in Automated 3-D Breast Ultrasound Using Coronal Spiculation, IEEE Trans. Med. Imag. 31 (2012) 1034–1042. D. Kotsianos-Hermle, S. Wirth, T. Fischer, K.M. Hiltawsky, M. Reiser, First clinical use of a standardized three-dimensional ultrasound for breast imaging, Eur. J. Radiol. 71 (2009) 102–108. D. Kotsianos-Hermle, K.M. Hiltawsky, S. Wirth, T. Fischer, K. Friese, M. Reiser, Analysis of 107 breast lesions with automated 3D ultrasound and comparison with mammography and manual ultrasound, Eur. J. Radiol. 71 (2009) 109–115. S.P. Sinha, M.A. Roubidoux, M.A. Helvie, A.V. Nees, M.M. Goodsitt, G.L. LeCarpentier, J.B. Fowlkes, C.L. Chalek, P.L. Carson, Multi-modality 3D breast imaging with X-ray tomosynthesis and automated ultrasound. in: Proceedings of the 29th Annual International Conference of the IEEE EMBS Cite Internationate, Lyon, France, August, 2007, pp. 23–26. X. Lin, J.W. Wang, F. Han, J.H. Fu, A.H. Li, Analysis of eighty-one cases with breast lesions using automated breast volume scanner and comparison with handheld ultrasound, Eur. J. Radiol. 81 (2012) 873–878. K.M. Kelly, J. Dean, W.S. Comulada, S.J. Lee, Breast cancer detection using automated whole breast ultrasound and mammography in radiographically dense breasts, Eur. Radiol. 20 (2010) 734–742. K.M. Kelly, J. Dean, S.J. Lee, W.S. Comulada, Breast cancer detection: radiologists’ performance using mammography with and without automated whole-breast ultrasound, Eur. Radiol. 20 (2010) 2557–2564. K.K. Shung, The principle of multidimensional arrays, Eur. J. Echocardiogr. 3 (2002) 149–153. T.R. Nelson, D.H. Pretorius, Three-dimensional ultrasound imaging, Ultrasound Med. Biol. 24 (1998) 1243–1270. M. Rao, T. Varghese, Correlation analysis of three-dimensional strain imaging using ultrasound two-dimensional array transducers, J. Acoust. Soc. Am. 124 (2008) 1858–1865. W.W. Jiang, A.H. Li, Y.P. Zheng, A semi-automated 3-D annotation method for breast ultrasound imaging: system development and feasibility study on phantoms, Ultrasound Med. Biol. 40 (2014) 434–446. C.D. Barry, C.P. Allott, N.W. John, P.M. Mellor, P.A. Arundel, D.S. Thomson, J.C. Waterton, Three-dimensional freehand ultrasound: image reconstruction and volume analysis, Ultrasound Med. Biol. 23 (1997) 1209–1224. Q.H. Huang, Y.P. Zheng, A new scanning approach for limb extremities using a water bag in freehand 3-D ultrasound, Ultrasound Med. Biol. 31 (2005) 575– 583. R.W. Prager, R.N. Rohling, A.H. Gee, L. Berman, Rapid calibration of 3-D freehand ultrasound, Ultrasound Med. Biol. 24 (1998) 855–869. A.K. Thittai, J.M. Yamal, J. Ophir, Small breast lesion classification performance using the normalized axial-shear strain area feature, Ultrasound Med. Biol. 39 (2013) 543–548. R. Prager, A. Gee, L. Berman, Stradx: real-time acquisition and visualization of freehand three-dimensional ultrasound, Med. Image Anal. 3 (1999) 129–140.
[35] G. Treece, R. Prager, A. Gee, L. Berman, Fast surface and volume estimation from non-parallel cross-sections, for freehand 3-D ultrasound, Med. Image Anal. 3 (1999) 141–173. [36] A. Gee, R. Prager, G. Treece, L. Berman, Narrow-band volume rendering for freehand 3D ultrasound, Comput Graph. 26 (2002) 463–476. [37] A. Fenster, D.B. Downey, 3-D ultrasound imaging: a review, IEEE Eng. Med. Biol. Mag. 15 (1996) 41–51. [38] O.V. Solberg, F. Lindseth, H. Torp, R.E. Blake, T.A.N. Hernes, Freehand 3D ultrasound reconstruction algorithms – a review, Ultrasound Med. Biol. 33 (2007) 991–1009. [39] Q.H. Huang, Y.P. Zheng, M.H. Lu, Z.R. Chi, Development of a portable 3D ultrasound imaging system for musculoskeletal tissues, Ultrasonics 43 (2005) 153–163. [40] Q.H. Huang, Y.P. Zheng, An adaptive squared-distance-weighted interpolation for volume reconstruction in 3D freehand ultrasound, Ultrasonics 44 (2006) e73–e77. [41] J.L. Gennisson, G. Cloutier, Sol-gel transition in agar-gelatin mixtures studied with transient elastography, IEEE Trans. Ultrason., Ferroelectr., Freq. Control 53 (2006) 716–723. [42] T.M. Mak, Y.P. Huang, Y.P. Zheng, Liver fibrosis assessment using transient elastography guided with real-time B-mode ultrasound imaging: a feasibility study, Ultrasound Med. Biol. 39 (2013) 956–966. [43] Y.P. Huang, Y.P. Zheng, S.F. Leung, A.F.T. Mak, Reliability of measurement of skin ultrasonic properties in vivo: a potential technique for assessing irradiated skin, Skin Res. Technol. 13 (2007) 55–61. [44] G. Rankin, M. Stokes, Reliability of assessment tools in rehabilitation: an illustration of appropriate statistical analyses, Clin. Rehab. 12 (1998) 187–199. [45] P.E. Shrout, J.L. Fleiss, Intraclass correlations: uses in assessing rater reliability, Psychol. Bull. 86 (1979) 420–428. [46] A.T. Stavros, Breast Ultrasound, Lippincott Williams & Wilkins, Philadelphia, 2004. pp. 46–47. [47] A.M. Dixon, Breast Ultrasound: How, Why and When, Churchill Livingstone, Edinburgh, 2008. pp. 94–99. [48] Z. Xue, K.K. Loo, J. Cosmas, M. Tun, L. Feng, P.Y. Yip, Error-resilient scheme for wavelet video codec using automatic ROI detection and Wyner–Ziv coding over packet erasure channel, IEEE Trans. Broadcast. 56 (2010) 481–493. [49] S.J. Ye, Y.F. Gu, Y. Zhan, An automatic ROI coding algorithm based on multiresolution target detection, in: 2008 2nd International Symposium on Systems and Control in Aerospace and Astronautics, vol. 1, issue: 2, 2008, pp. 610–613. [50] C.K. Modi, N.P. Desai, A simple and novel algorithm for automatic selection of ROI for dental radiograph segmentation. in: 24th Canadian Conference on Electrical and Computer Engineering, Niagara Falls, Canada, May 08–11, 2011. [51] M.J. Stoutjesdijk, J. Veltman, H. Huisman, N. Karssemeijer, J.O. Barentsz, J.G. Blickman, C. Boetes, Automated analysis of contrast enhancement in breast MRI lesions using mean shift clustering for ROI selection, J. Magn. Reson. Imag. 26 (2007) 606–614. [52] T.D. Nguyen, S.H. Kim, N.C. Kim, An automatic body ROI determination for 3D visualization of a fetal ultrasound volume, in: 9th International Conference on Knowledge-Based Intelligent Information and Engineering Systems, La Trobe University, Melbourne, Australia, September 14–16, 2005. [53] Y. Su, Y. Wang, J. Jiao, Y. Guo, Automatic detection and classification of breast tumors in ultrasonic images using texture and morphological features, Open Med. Inform. J. S1–3 (2011) 26–37. [54] W.R. Crum, T. Hartkens, D.L.G. Hill, Non-rigid image registration: theory and practice, Br. J. Radiol. 77 (2004) S140–S153. [55] Y.J. Guo, R. Sivaramakrishna, C.C. Lu, J.S. Suri, S. Laxminarayan, Breast image registration techniques: a survey, Med. Biol. Eng. Comput. 44 (2006) 15–26. [56] C. Leung, K. Hashtrudi-Zaad, P. Foroughi, P. Abolmaesumi, A real-time intrasubject elastic registration algorithm for dynamic 2-D ultrasound images, Ultrasound Med. Biol. 35 (2009) 1159–1176. [57] J.F. Krucker, G.L. LeCarpentier, J.B. Fowlkes, P.L. Carsen, Rapid elastic image registration for 3-D ultrasound, IEEE Trans. Med. Imag. 21 (2002) 1384–1394. [58] G.F. Xiao, J.M. Brady, J.A. Noble, M. Burcher, R. English, Nonrigid registration of 3-D free-hand ultrasound images of the breast, IEEE Trans. Med. Imag. 21 (2002) 405–412. [59] S. Khallaghi, C.G.M. Leung, K. Hastrudi-Zaad, P. Foroughi, C. Nguan, P. Abolmaesumi, Experimental validation of an intrasubject elastic registration algorithm for dynamic-3D ultrasound images, Med. Phys. 39 (2012) 5488– 5497. [60] A.H. Gee, G.M. Treece, R.W. Prager, C.J.C. Cash, L. Berman, Rapid registration for wide field of view freehand three-dimensional ultrasound, IEEE Trans. Med. Imag. 22 (2003) 1344–1357. [61] G.M. Treece, R.W. Prager, A.H. Gee, L. Berman, Correction of probe pressure artifacts in freehand 3D ultrasound, Med. Image Anal. 6 (2002) 199–214.