An Algorithm for the Pulmonary Border Extraction in PET Images

Available online at www.sciencedirect.com

Procedia Technology 5 (2012) 876 – 884

CENTERIS 2012 - Conference on ENTERprise Information Systems / HCIST 2012 - International Conference on Health and Social Care Information Systems and Technologies

An Algorithm for the Pulmonary Border Extraction in PET Images Fábio Nerya,b,c, José Silvestre Silvad,e*, Nuno C. Ferreirab,c, Francisco Caramelob and Ricardo Faustinob,c a Department of Physics, Faculty of Sciences and Technology, University of Coimbra, Portugal IBILI – Institute of Biomedical Research in Light and Image, Faculty of Medicine, University of Coimbra, Portugal c ICNAS – Institute of Nuclear Sciences Applied to Health, University of Coimbra, Portugal d Instrumentation Center, Faculty of Sciences and Technology, University of Coimbra, Portugal e School of Technology and Management, Polytechnic Institute of Portalegre, Portugal

b

Abstract Automatic segmentation methods are mandatory for most computer-aided diagnostic methods, particularly in Positron Emission Tomography (PET), being extremely relevant for quantification purposes, disease diagnosis and staging. These automated approaches speed up the clinical work-flow, eliminating the need of a tedious manual delineation by physicians and greatly improving the reproducibility of the delineation procedures. This paper presents a novel approach for the fully automatic extraction of the pulmonary boundaries for PET images based in the concept of "marker-driven watershed segmentation". Additionally, an algorithm for the lung border extraction in CT images was developed in response to physicians' requirements for a better understanding of each individuals' specific anatomy. The accuracy of both algorithms was assessed, comparing the results of both approaches to their correspondent manually depicted contours by several physicians, taking into account several figures of merit. © 2012 Published by Elsevier Ltd. Selection and/or peer review under responsibility of CENTERIS/SCIKA -

© 2012 Published by Elsevier Ltd. Selection and/or peer-review under responsibility of CENTERIS/HCIST. Association for Promotion and Dissemination of Scientific Knowledge Keywords: Image Segmentation; Watershed Segmentation; Positron Emission Tomography; Computed Tomography; Pulmonary Imaging

* Corresponding author. Tel.: +351-239-410-600. E-mail address: [email protected].

2212-0173 © 2012 Published by Elsevier Ltd. Selection and/or peer review under responsibility of CENTERIS/SCIKA - Association for Promotion and Dissemination of Scientific Knowledge doi:10.1016/j.protcy.2012.09.097

Fábio Nery et al. / Procedia Technology 5 (2012) 876 – 884

1. Introduction Positron Emission Tomography (PET) is a nuclear medicine imaging technique that produces images conveying information regarding the functional processes of the organism. It is particularly important and widely used in Oncology, where in combination with several radiotracers, such as the glucose analog 18F-FDG (fluorodeoxyglucose), allows the exploration of the distinct biochemical behavior between normal and neoplastic tissue. The ever-increasing size of the datasets supplied by current medical imaging techniques leads to the need for automatic methods for the analysis of all the available information. Image segmentation methods play a significant role in the development of these methods and also when applied to PET images for the wide range of body structures it is essential for the study of several phenomena, such as the interference of motion during the acquisition stage (particularly important for the lungs), image artifacts and the influence of different acquisition protocols. Nevertheless, the development of image segmentation algorithms for PET images is a challenging task given their poor intrinsic spatial resolution (comparatively to anatomical modalities such as CT and MRI) and significant noise component. The present work is a fully automatic method for the segmentation of the lungs in PET images based in the concept of “marker-driven watershed segmentation” (see Fig. 2). A close contact was kept with several physicians during the development of this approach and since they frequently resort to CT images to obtain a better understanding of each patients' specific anatomy, we also simultaneously developed a lung segmentation algorithm for CT images (see Fig. 3). 2. State of the Art 2.1. Positron Emission Tomography Some approaches have been used so far for PET volume delineation, particularly for the segmentation of tumors. A common approach are thresholding methods [1], where occasionally connected component labelling procedures are also performed [2]. Support Vector Machines (SVM) can be used to obtain the optimal threshold value [3]. Additionally, other approaches include partial differential equation (PDE)-based methods, both in semi-automatic [4] and automatic [5] procedures. Furthermore, the use of a priori knowledge using computed tomography (CT) has also shown to improve results of the delineation of PET lung tumor images [6]. A modified active contour model, named Poisson Gradient Vector Flow (PGVF), was used for the automatic segmentation of the liver in PET images [7]. The initial contour is defined through the application of a Canny edge detector where its parameters (σ, threshold value) are determined using a genetic algorithm. Tylski et al. [8] developed an interactive watershed segmentation method for the delineation of simulated tumors using phantom images. This method requires the user to place markers in the object to be segmented and in the background. 2.2. Computed Tomography Lung segmentation in CT, unlike in PET, is widely performed where a broad range of distinct methods have already been implemented [9]. In addition to the common thresholding approaches [10-14], several other methods have been applied, such as region growing approaches [15], methods based on level sets [16, 17], clustering approaches [18], active shape models such as [19], where the first position of the active shape model is determined by a rib detection method, adaptive border marching approaches [20], taking into account the juxtapleural nodules, among many others. Several of these approaches were also used to segment other regions such as heart cavities [21, 22] and vascular or neural structures [23, 24]. These methods can assist the physicians in the follow-up of pulmonary pathologies [25, 26].

877

878

Fábio Nery et al. / Procedia Technology 5 (2012) 876 – 884

3. Segmentation Methods 3.1. Positron Emission Tomography x Watershed Segmentation – Theory: The concept of watersheds is based on the visualization of a graylevel image as a topographic surface where the two spatial coordinates in addition to the intensity of the image make the 3D space. The basic idea is illustrated in Fig. 1, which is to flood the surface from below (regional minima) and when the water rising from two distinct catchment basins is about to merge, a dam is constructed in order to prevent the merging. Eventually, only the tops of the dams will be exposed above the water and they will correspond to the connected boundaries extracted by the algorithm.

Fig. 1. Illustration of the flooding process (a)-b)) and dam construction (c)-d))

x Algorithm Description: The first step of the algorithm is the noise attenuation using a Wiener adaptive filter. The computation of the morphological watershed in a gradient image leads to an oversegmentation, yielding unusable results, due to several factors such as noise and other irregularities of the gradient image. In order to prevent this, our method automatically introduces a priori knowledge through the use of automatic markers that “control” the segmentation process. There are two different kinds of markers: internal markers, that acting as seeds for the segmentation procedure are required to be located inside the objects of interest; and external markers, which are part of the background (see Fig. 2). In order to define the set of objects corresponding to the internal markers, we first apply of a double-threshold to the filtered image to create two different masks: one containing all pixels belonging to the body of the patient and the other containing “low uptake structures”, where the lungs are included. We obtain the line representing the “border” of the body and perform a morphological dilation a specific number of times depending on the distance between the “border” of the body and the largest “low uptake structures”. This step allows us to eliminate structures whose uptake is similar to the lungs but are too close to the skin surface. In terms of the external markers' definition, we found that two different types of markers were needed. The first type is a line behaving as a barrier preventing an excessive growth of the region under the segmentation process and the consequent “leaking” to regions outside of the patient, since there is no uptake there. We define this line by performing a new dilation of the patients' body border and extracting the inner contour of the resulting object. The second type of external marker is also used to prevent an excessive growth that, in this case, may result in the merging of the two lungs. The line that vertically divides the two sides of the image is kept and used as an external marker. The next step is the minima imposition in two images: the Wiener filtered image and its smoothed gradient. By imposing the new minima in both images, we modify their intensities so that their regional minima are at the location of the markers. The watershed segmentation of both modified images is performed and the results are intersected, yielding the results shown in Fig. 2.

Fábio Nery et al. / Procedia Technology 5 (2012) 876 – 884

Fig. 2. Watershed segmentation results for two different PET slices of two different patients. Red: Internal Markers; Light Blue and Yellow: External Markers; Green: Final Result. The lung borders were successfully computed. Notice the importance of the external markers in definition of the correct borders, particularly if the determination of the internal markers isn't fully accurate (a))

3.2. Computed Tomography The algorithm starts with a filtering operation with the goal of noise reduction [27], through the application a Gaussian filter. The next step is the threshold operation where the optimal value is computed using information extracted from the volumes' intensity histogram by searching for the local minimum. Connected component labelling operations are performed subsequently in order to remove regions outside of the body of the patient and all slices outside of the pulmonary region are rejected. In order to improve the segmentation results, the large airways are removed by locating the trachea in the first slice of the pulmonary region, removing it. A search for the airways, in subsequent slices, is performed, within the referred region of interest, which is updated for each slice. The algorithm is able to detect if the airways are merged with the lung area, selectively removing only the airways, which are identified through a new specific threshold procedure and morphological processing. Additionally, if the merging of the lungs is verified, the combination of new morphological operations and an iterative threshold results in the identification of the pixels that make up the thin area between the lungs, which then are removed. Fig. 3 shows the sequence of operations performed in the lung separation algorithm (b)-h)) and the final result of the proposed segmentation approach (i)). 4. Validation of the Algorithms During the development of the algorithm, we required the help of an expert physician with several years of experience to frequently evaluate the effectiveness of the method, guiding our efforts in order to obtain an accurate delineation of the lung borders. Given the necessity of a quantitative evaluation for a thorough validation [28] of the proposed approach, we asked several physicians to manually draw the pulmonary contours of several slices, using a graphical user interface (GUI) that we designed specifically for this purpose. For the validation of the PET segmentation algorithm, we had the collaboration of two expert physicians that performed a manual delineation of 30 slices with dimension of 144x144 pixels (corresponding to 60 lung borders). The validation of the CT segmentation algorithm was performed using reference contours traced by 3 expert observers in 19 different CT slices. Two of the three observers have drawn the same contours two times, with an elapsed time between each delineation of about one month. In order to compare the lung borders computed by our algorithm with the manually drawn contours, several figures of merit were used, as described in section 4.1.

879

880

Fábio Nery et al. / Procedia Technology 5 (2012) 876 – 884

Fig. 3. Algorithm to detach merged lungs; a) Merged lungs and airways; b) Merging area; c) Upper and lower borders of b); d) Extreme opposite pixels; e)-f) Dynamic threshold; g) Extreme pixels of the area of separation; h) Area to remove; i) Final result of the algorithm

4.1. Figures of Merit x Pratt Figure of Merit: The Pratt Figure of Merit (FPRATT) [28] is defined as:

FPRATT

1 N 1 , ¦ N i 1 1  D u di2

(1)

where N is the total number of points, α is a normalization parameter and di is the distance between correspondent points. Therefore, a value of 1 for FPRATT indicates total overlap between two contours. x Mean Error: The mean error (Dmean) is simply the mean of the distances between corresponding points (di) as defined in the following expression:

Dmean

1 N 2 ¦ di N i1

5. Results and Discussion 5.1. PET Segmentation Algorithm The results for the mean error of the comparison of the output of our algorithm for the 30 slices and the correspondent slices drawn by the two experts are depicted on Fig. 4, where the labelling convention is: AAlgorithm; Ph-Human Observer; T-Time Instance. As we can see, the variability between the contours that our approach produces in comparison with the contours drawn by the physicians is comparable to the variability of contours drawn by different experts. Fig. 5 presents the Bland-Altman plots comparing the areas

(2)

Fábio Nery et al. / Procedia Technology 5 (2012) 876 – 884

confined within the borders computed by our method and the average of the areas of the manually traced contours. Once more, the area differences between the contours of both observers are examined. In most cases, our approach tends to overestimate the lung areas comparatively to the human observers being the mean difference of areas of about 100 pixels. Nevertheless, given the intrinsic nature of the technique which creates difficulties in the development of very-high accuracy segmentation methods, we consider that the errors performed by our approach are acceptable.

Fig. 4. Box plot for the mean error of the comparison of the lung borders computed by our algorithm and the manually delineated borders. Additionally, the inter-observer variability is also assessed

Fig. 5. Bland-Altman plots for the comparison of the areas enclosed by the borders computed by our lung segmentation algorithm for PET images and the average of the correspondent areas of the borders drawn by expert observers

5.2. CT Segmentation Algorithm A study of the variability between the algorithm and the observers, inter-observer variability and intraobserver variability, for 2 different experts was performed. The results are depicted in Fig. 6 for the Pratt figure of merit. Regarding the variability between the comparisons, once more we can conclude that the “algorithm-observer” variability is comparable (even better in some cases) to the inter-observer and intraobserver variability. To enhance the validation set, one of the experts has drawn several additional slices of

881

882

Fábio Nery et al. / Procedia Technology 5 (2012) 876 – 884

additional patients, totalling 41 slices (82 lung contours) of 6 different patients. Fig. 7 depicts the BlandAltman plot obtained through the comparison of the areas resulting from the CT segmentation algorithm and the manual approach. We can conclude that the algorithm neither has the tendency to over or underestimate the lung areas, being the average difference of correspondent contours very close to zero. Additionally, the fact that the area difference values are small compared to average area values also indicates the effectiveness of the method.

Fig. 6. Boxplots of the Pratt Figure of Merit for the comparison of the algorithm with 3 different observers (2 of which drawn the same contours in separate time instances). Additionally, the inter-observer and intra-observer comparisons are also presented

Fig. 7. Bland-Altman plot obtained from the area measurements of the CT lung borders computed by our approach and by a manual delineation performed by expert observers using 41 CT slices (corresponding to 82 lung borders)

Fábio Nery et al. / Procedia Technology 5 (2012) 876 – 884

6. Future Work Our future goals are not only to keep improving our algorithms, particularly through the development of a 3D approach for the PET segmentation but also to conduct new validation tests using a dataset comprising a greater number of exams. We would also like to develop an algorithm to perform the fusion of the distinct lung volumes that both imaging modalities provide which may help to identify areas where the PET/CT registration is not fully accurate. Additionally, these methods can also help to identify possible attenuation correction errors that may result in some artifacts in the PET images.

Acknowledgements The authors would like to thank ICNAS for providing the PET and CT exams and to the experts that collaborated in the validation of both algorithms.

References [1] [2]

[3]

[4]

[5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16]

[17]

Jentzen W, Freudenberg L, Eising EG, Heinze M, Brandau W, Bockisch A. Segmentation of PET Volumes by Iterative Image Thresholding. J. Nucl. Med. 2007; 48:108-114 Kanakatte A, Gubbi J, Mani N, Kron T, Binns D. A Pilot Study of Automatic Lung Tumor Segmentation from Positron Emission Tomography Images using Standard Uptake Values. IEEE Symposium on Computational Intelligence in Image and Signal Processing – CIISP 2007, IEEE Press, pp. 363-368 Kerhet A, Small C, Quon H, Riauka T, Greiner R, McEwan A, Roa W. Segmentation of Lung Tumours in Positron Emission Tomography Scans: a Machine Learning Approach. Artificial Intelligence in Medicine, Lecture Notes in Computer Science 2009; 5651:146-155 Kanakatte A, Gubbi J, Mani N, Kron T, Binns D, Palaniswami M. Pulmonary Tumor Volume Delineation in PET Images using Deformable Models. 30th Annual International IEEE Engineering in Medicine and Biology Society Conference 2008, pp. 31183121 Dedić R, Allili M, Lecomte R. Unsupervised Cardiac PET Image Segmentation. 18th IEEE International Conference on Systems, Signals and Image Processing – IWSSIP 2011, pp. 273-278 Kim J, Wen L, Eberl S, Fulton R, Feng D. Use of Anatomical Priors in the Segmentation of PET Lung Tumor Images. IEEE Nucl. Sci. Symp. Conf. Record 2007. Vol.6, pp. 4242-4245 Hsu C, Liu C, Chen C. Automatic Segmentation of Liver PET Images. Computerized Medical Imaging and Graphics 2008;32:601610 Tylski P, Bonniaud G, Decencière E, Stawiaski J, Coulot J, Lefkopoulos D, Ricard M. 18F-FDG PET Images Segmentation using Morphological Watershed: A Phantom Study, Proc. IEEE Nucl. Sci. Symp. Conf. Record 2006, Vol. 4, pp. 2063-2067 Sluimer I, Schilham A, Prokop M, Ginneken B. Computer Analysis of Computed Tomography Scans of the Lung: A Survey. IEEE Trans. Med. Imag. 2006; 25(4):385-405 Hu S, Hoffman E, Reinhardt J: Automatic Lung Segmentation for Accurate Quantitation of Volumetric X-ray CT Images. IEEE Trans. Med. Imag. 2001; 20(6):490-498 Gao Q, Wang S, Zhao D, Liu J. Accurate Lung Segmentation For X-ray CT Images. IEEE Third International Conference on Natural Computation 2007, Vol.2, pp. 275-279 Tseng L, Huang L: An Adaptive Thresholding Method for Automatic Lung Segmentation in CT Images. IEEE AFRICON 2009, pp. 1-5 Silva JS, Silva A, Santos BS, Madeira J. Detection and 3D Representation of Pulmonary Air Bubbles in HRCT Volumes. SPIE Medical Imaging 2003: Physiology and Function: Methods, Systems, and Applications, Vol. 5031, pp. 430-439 Silva JS, Silva A, Santos BS. A Volumetric Pulmonary CT Segmentation Method with Applications in Emphysema Assessment. SPIE Medical Imaging 2006: Physiology, Function, and Structure from Medical Images, Vol. 6143, pp. 885-896 Sun X, Zhang H, Duan H. 3D Computerized Segmentation of Lung Volume with Computed Tomography. Academic Radiol. 2006; 13(6):670-677 Silva JS, Santos BS, Silva A. A Level-Set Based Volumetric CT Segmentation Technique: A Case Study with Pulmonary Air Bubbles. International Conference on Image Analysis and Recognition - ICIAR 2004. Lecture Notes in Computer Science, Vol. 3212, pp. 68-75 Silveira M, Nascimento J, Marques J. Automatic Segmentation of the Lungs using Robust Level Sets. Proc. IEEE EMBS Conf. 2007, pp. 4414-4417

883

884

Fábio Nery et al. / Procedia Technology 5 (2012) 876 – 884

[18] Jaffar M, Hussain A, Mirza A. Fuzzy Entropy Based Optimization of Clusters for the Segmentation of Lungs in CT Scanned Images. Knowledge and Information Systems 2012; 24(1):91-111 [19] Sun S, Bauer C, Beichel R. Automated 3-D Segmentation of Lungs With Lung Cancer in CT Data Using a Novel Robust Active Shape Model Approach. IEEE Trans. Med. Imag. 2012; 31(2):449-460 [20] Pu J, Roos J, Yi C, Napel S, Rubin G, Paik D. Adaptive Border Marching Algorithm: Automatic Lung Segmentation on Chest CT Images. Computerized Medical Imaging and Graphics 2008; 32(6): 452-462 [21] Antunes SG, Silva JS, Santos JB, Martins P, Castela E. Phase Symmetry Approach Applied to Children Heart Chambers Segmentation: A Comparative Study, IEEE Trans. Biomed. Eng. 2011; 58(8):2264-2271. [22] Antunes SG, Silva JS, Santos JB. A Level Set Segmentation Method of the Four Heart Cavities in Pediatric Ultrasound Images. International Conference on Image Analysis and Recognition – ICIAR 2010. Lecture Notes in Computer Science vol. 6112, part II, pp. 99 – 107. [23] Ferreira A, Morgado AM, Silva JS. A Method for Corneal Nerves Automatic Segmentation and Morphometric Analysis. Computer Methods and Programs in Biomedicine 2012; 107:53-60. [24] Ferreira A, Morgado AM, Silva JS. Automatic Corneal Nerves Recognition for Earlier Diagnosis and Follow-up of Diabetic Neuropathy. Lecture Notes in Computer Science 2010; 6112, part II: 60-69. [25] Ferreira C, Santos BS, Silva JS, Silva A. Comparison of a Segmentation Algorithm to Six Expert Imagiologists in Detecting Pulmonary Contours on X-ray CT Images. SPIE Medical Imaging 2003: Image Perception, Observer Performance and Technology Assessment 2003, Vol. 5034, pp. 347-358 [26] Silva J., Cancela J, Teixeira L. Fast Volumetric Registration Method for Tumor Follow-Up in Pulmonary CT Exams. Journal of Applied Clinical Medical Physics 2011; 12: 362-375 [27] Silva JS, Silva A, Santos BS. Image Denoising Methods for Tumor Discrimination in High Resolution Computed Tomography. Journal of Digital Imaging 2011; 24:464-469 [28] Santos B, Ferreira C., Silva JS, Silva A, Teixeira L. Quantitative Evaluation of a Pulmonary Contour Segmentation Algorithm in Xray Computed Tomography Images. Academic Radiol. 2004; 11(8):868-878