Left Ventricle Segmentation from Heart MDCT

Transcription

1 Left Ventricle Segmentation from Heart MDCT Samuel Silva, Beatriz Sousa Santos, Joaquim Madeira, and Augusto Silva DETI / IEETA University of Aveiro, Aveiro, Portugal sss@ua.pt Abstract. A semi-automatic method for left ventricle segmentation from MDCT exams is presented. It was developed using ITK and custom modules integrated in the MeVisLab platform. A preliminary qualitative evaluation shows that the provided segmentation, without any tweaking or manual edition, is reasonably close to the ideal segmentation as judged by experienced radiology technicians. 1 Introduction Multi-detector row computer tomography (MDCT) systems have been evolving at a fast pace enabling submillimeter spatial resolutions and low exam times, allowing full heart imaging in less than a breath-hold and providing 10+ heart volumes per cardiac cycle. The same data obtained for coronary angiography can be used to perform analysis of left ventricle function and myocardial perfusion with no need of extra radiation exposure, leading to additional information which can help the practitioner to attain a more complete diagnosis. In fact, recent studies (e.g., [1]) comparing data obtained from different imaging modalities, reveal that reliable left ventricle analysis can be performed using MDCT data. As stated by Cury et al. [2], the full potential of MDCT technology is just beginning to be explored and further advances are essential, concerning not only new automatically computed parameters but also improved tools to enable further insight into the wide range of data now becoming available (e.g., for left ventricle wall motion). At the ground level of such developments are segmentation methods. Different approaches regarding left ventricle segmentation for a variety of exam modalities have been widely described in the literature in the past years [3] and range from 2D methods (e.g., [4]) to more complex, model-based, approaches (e.g., [5]). Unfortunately, to the best of our knowledge, no off-the-shelf segmentation method is available which could be used as a basis for our work. Therefore, we engaged in the development of a segmentation method which should: 1) be as automatic as possible; 2) segment both left ventricle endocardium and epicardium; 3) provide a segmentation for all cardiac phases included in the MDCT exam; and 4) be validated by radiology technicians to ensure the reliability of the resulting data. This article describes a semi-automatic left ventricle segmentation method which was implemented using ITK ( and custom processing modules integrated in the MeVisLab ( H. Araujo et al. (Eds.): IbPRIA 2009, LNCS 5524, pp , c Springer-Verlag Berlin Heidelberg 2009

2 Left Ventricle Segmentation from Heart MDCT 307 platform. The results from a qualitative evaluation of the segmentation results, conducted with the collaboration of experienced radiology technicians, are presented and show that the resulting endocardium segmentation (without any parameter tweaking or manual edition) is of good quality, requiring, in general, minimum edition, while the epicardium segmentation performs reasonably but clearly requires further enhancements. In the next sections the implemented segmentation method is presented followed by a description of the evaluation process and obtained results. Finally, we present some ideas for future work. 2 Left Ventricle Segmentation Coronary angiography MDCT exams are used (i.e., with contrast agent present [6], mainly in the coronary arteries and left ventricle), containing 10 to 12 heart volumes, along one cardiac cycle, each having a resolution of voxels. To improve image quality before segmentation, two methods are used. First, a minimum intensity filter is applied along the YY axis using a three voxel neighborhood. Then, to partially attenuate noise, a smoothing filter is applied. The developed segmentation method has four main stages: 1) left ventricle principal axis estimation; 2) reference phase endocardium segmentation; 3) full exam endocardium segmentation; 4) full exam epicardium segmentation. The main aspects of each stage will be explained in the following sections. 2.1 Principal Axis Estimation Left ventricle axes must be estimated in order to select a proper orientation for the data and segment the left ventricle using slices along its principal axis. From the 10 to 12 cardiac phases available in each exam, the one corresponding to 60% of the cardiac cycle was chosen for left ventricle principal axis estimation, since it exhibited better image quality. This is due to the fact that it belongs to the diastolic phase of the cardiac cycle where structural motion is less significant and a higher amount of radiation is applied to ensure better image quality. During the remaining phases of the exam the radiation levels are kept to a minimum (to reduce radiation dosage [6]) resulting in images with worse quality. To estimate left ventricle position a coronal slice at midventricular level is analysed. A threshold is performed and the image is searched from right to left looking for active regions. The first region found with a large area (to exclude regions related with the coronary arteries, lungs, etc.) is considered part of the left ventricle. Region growing is then applied in order to isolate it from all other active regions on the image (including, eventually, portions of the right ventricle where some contrast agent might be present). Finally, the centroid of the segmented region is computed and stored to be used as a reference during the segmentation of the remaining slices. In order to ease the detection of slices close to the left ventricle apex (where area cannot be used as a detection criteria), 3D region growing is applied to the

3 308 S. Silva et al. Fig. 1. Slice showing the segmented endocardium not including the papillary muscles (left) and including them in the blood pool (right) entire set using the previously computed centroid as seed. This will isolate the left ventricle blood pool from other active regions related with the lungs and ribs which might result in classifying a region as part of the left ventricle too early (ribs) or out of place (lungs). The data set resulting from the previously performed 3D region growing is then processed, slice by slice, along the coronal plane, from the left ventricle apex to its base. Each slice is searched for an active region from right to left. To ensure that none of the possibly remaining regions are misclassified as left ventricle the one closest to the previously computed reference centroid is chosen. It follows region growing to segment that region and hole filling to ensure that holes related with the papillary muscles are closed. To detect the stopping slice, the leftmost pixel of each segmented region is computed. If a sudden decrease (i.e., it moves to the left of the image) occurs it means that the region where the left ventricle connects with the aorta has been reached (outflow tract). From that moment on, our method looks for a sudden increase in the leftmost pixel of the segmented region which is related with the entrance in the left atrium region and segmentation stops. Finally, the centroids for the first and last slices are used to compute the left ventricle principal axis. The exam is then rotated in order to position the left ventricle axis normally to the coronal plane and the remaining processing is performed using this orientation. 2.2 Reference Phase Segmentation Starting from the 60% phase (due to better image quality), the endocardium is segmented using a method similar to that used to estimate the principal axis. The blood pool centroid and maximum radius are computed at a midventricular slice and 3D region growing is applied using that voxel as seed. The resulting data set is processed, slice by slice, along the coronal plane, from the left ventricle apex to its base. Each slice is searched for an active region from right to left. To ensure that none of the possibly remaining regions are misclassified as left ventricle the one closest to the previously computed reference centroid is chosen. It follows region growing and hole filling to include the papillary muscles in the blood pool (see figure 1).

4 Left Ventricle Segmentation from Heart MDCT 309 (a) (b) Fig. 2. Segmented endocardium (a) and epicardium (b) for end-systolic and enddiastolic phases The stopping slice (i.e., one desirably at mitral valve level) is detected by a significant change of blood pool radius, when compared with the reference radius obtained earlier (entrance in the outflow tract and left atrium). 2.3 Full Exam Segmentation After segmenting the endocardium for the reference phase it followsendocardium and epicardium segmentation for all exam phases. Endocardium. The reference endocardium is scaled (to encompass a slightly larger blood volume for the end-diastolic phase) and used as a mask to isolate the left ventricle blood pool in all exam phases. Next, thresholding is applied followed by hole filling. Due to left ventricle motion the location of the mitral valve changes. Thus, further processing is necessary to estimate that plane for each phase. A reference blood pool radius is computed for a midventricular slice and the segmentation stopped when blood pool radius significantly exceeds that value (entrance in the left atrium and outflow tract). Figure 2(a) shows two segmentations obtained for end-sistole and end-diastole. Epicardium. Epicardium segmentation is harder to perform due to poor image quality, including a slight variation of the greylevel interval corresponding to the epicardium along the different exam phases, and due to the existence of pixels inside that interval in adjacent regions such as the right ventricle and liver. To minimize the effects arising from these issues, left ventricle radius is estimated for each phase in order to define a region of interest as narrow as possible and pixels are sampled along the epicardium to help define a suitable greylevel interval for segmentation using thresholding. Figure 2(b) shows two segmented epicardia for end-sistole and end-diastole. 3 Evaluation At this development stage we opted for a less-costly, preliminary qualitative evaluation that would allow the detection of any serious segmentation problem and inform further fine tuning of the proposed algorithm.

5 310 S. Silva et al. (a) (b) Fig. 3. (a) Unedited segmentation results provided by a TeraRecon Aquarius workstation; and (b) unedited segmentation results, for the same exam, provided by our method Such a preliminary evaluation has been performed with the help of radiology technicians with 3 to 4 years daily experience in analysing MDCT heart images. 3.1 Methods Eventhough our method segments all exam phases, in this evaluation only three phases were evaluated per exam: the end-systole and end-diastole, which are used to compute important parameters such as the ejection fraction [1], and the 60% phase which was used as a reference for the segmentation of all remaining phases. An important aspect of every segmentation method is that the resulting first segmentation (i.e, without manual parameter tweaking or edition) should be as accurate as possible in order to minimize correction time. Having that in mind we decided to evaluate the segmentation results using only the standard values for the different parameters (e.g., ratio value between reference endocardium radius and current slice radius to determine mitral valve plane; determined empirically along the development) and no manual edition. These first segmentations were shown to the technicians and were considered good first approaches with no major abnormalities and comparable to first segmentations (specially for the endocardium) provided, for example, by a TeraRecon Aquarius workstation which additionally requires a priori manual alignment of the data in three different planes. Figure 3 shows an example where, even with the three plane initialization, the TeraRecon workstation provides a worst first segmentation. Notice the segmentation problems in the apex and midventricular slices. After such preliminary results we decided to perform a more thorough evaluation by asking the technicians to classify the resulting segmentations as if they were final. They were asked to evaluate segmentation results for a set of seven exams taken randomly from that week s exams. Care was only taken to ensure that no serious movement artifacts were present which might influence the results. Each technician received a grid to guide the evaluation process and where the evaluations should be registered. Regarding the endocardium, the grid considered segmentation evaluation for four anatomical regions deemed of interest: apex,

6 Left Ventricle Segmentation from Heart MDCT 311 midventricular slices, mitral valve and outflow tract. For each region the segmentation could be generically classified as OK (optimum segmentation), EXCESS (the segmented region is larger than the ideal) and SHORTAGE (the segmented region is smaller than the ideal). For the last two cases, the technician had to classify the severity of the problem using a three-level scale (1 low significance: the segmentation is very good, although, for the sake of perfectness, it could include/exclude a very small region; 2 moderate: the segmentation is good but it could be significantly improved by the inclusion/exclusion of a small region; 3 serious: the segmentation cannot be used without the inclusion/exclusion of important regions). For the epicardium a similar evaluation scale was used and five anatomical regions were considered: apex and the lateral (external) and septal (between ventricles) regions for the midventricular slices and base. Before evaluating the segmentations for each exam, the technicians were allowed to adjust cutting planes in order to define what they considered the proper 4-chamber (ventricles + atria) and left ventricle short-axis visualizations (which they usually use to analyse the left ventricle). Next, they were presented with those two visualizations of the different exam phases, with the segmentation superimposed onto them, and allowed to scroll along the slices parallel to the cutting planes (which could be changed, at any time, to encompass their needs). The opacity of the segmented regions could be changed in order, e.g., to better examine the regions bellow the segmentation mask. The endocardium and epicardium segmentations could be visualized separately or at the same time. The technicians could take all the time they needed to evaluate each phase and were allowed to alternate between phases for comparison (e.g., in order to assess if the papillary muscles were coherently included/excluded). 3.2 Results Table 1A shows evaluation results concerning endocardium segmentation. Notice that for most exams and cardiac phases the apex and midventricular slices where considered to be well segmented. Concerning the mitral valve and outflow tract there are different problems with different severities. A worst case example is exam 2 with level 3 problems in four situations. Nevertheless, notice that this just means a segmentation that must be edited before usage and not a completely abnormal segmentation. Figure 4(a) shows the endocardium segmentation for the end-sistole of exam 2 as an example of the worst segmentation found for the mitral valve region. Figure 4(b) shows the corresponding automatic segmentation as provided by a Siemens Circulation workstation with a significant region wrongly segmented inside the left atrium. Remarkably, during the evaluation, we noticed that two different criteria might be used to evaluate outflow tract segmentation: one considering that the tract should not be included in left ventricle segmentation and another considering that all the tract up to the aortic valve should be considered. This issue requires clarification and a deeper investigation is being carried out. Concerning epicardium segmentation (evaluation results presented in Table 1B), the technicians found most of the severe problems in the septal sections of both

7 312 S. Silva et al. Table 1. Evaluation results for endocardium and epicardium segmentation A. Endocardium Segmentation Evaluation End-Sistole End-Diastole 60% Exam Apex Mid. Valve Tract Apex Mid. Valve Tract Apex Mid. Valve Tract 1 O O - O O O O O O O + O 2 O O O O O O O O O O O ++ O O O O O + O O --- O O O -- O O O O O O O B. Epicardium Segmentation Evaluation End-Sistole End-Diastole Exam Apex Mid. Base Apex Mid. Base L S L S L S L S 1 O O +++ O +++ O O +++ O O O O O O O O 3 - O O O O O O - ++ O O O O O O:Optimum + : Excess : Shortage Number of symbols conveys severity level, e.g.: +++ : level 3 excess - - : level 2 shortage (a) (b) (c) Fig. 4. Endocardium segmentation, as provided by our nethod (a), for the end-distolic phase of exam 2 and corresponding segmentation as provided by a Siemens Circulation workstation (b). In (c), example of epicardium segmentation problem found in the septal section. midventricular and basal slices. Eventhough our method tries to adapt the threshold interval to each phase and seems to work very reasonably for some exams, the greylevels found in the septal section are, in many exam phases (due to noise and low radiation dosage) similar to those found inside the right ventricle. The initial mask applied to the image, to define a restricted region of interest, is limiting the severity of some of the problems but epicardium segmentation clearly needs further improvements. Figure 4(c) shows an example of a segmentation problem found in the septal section of the midventricular slices. The segmentation should roughly end in the dashed line. 4 Conclusions and Future Work A semi-automatic left ventricle segmentation method implemented using ITK and custom modules integrated in the MeVisLab platform is presented. A

8 Left Ventricle Segmentation from Heart MDCT 313 preliminary qualitative evaluation conducted using a set of seven exams evaluated by experienced radiology technicians provided interesting results: the endocardium segmentation (without any parameter tweaking or manual edition) exhibits good quality and epicardium segmentation, although still needing further enhancements concerning the segmentation of the septal section of the ventricle wall, yields promising results. Motivated by these results the following lines of work are considered relevant: Provide a simple editing tool to adjust the resulting segmentation which would easily solve the majority of the segmentation problems found; Improve the epicardium segmentation by using more robust statistics to characterize greylevel distribution (e.g., outlier removal), around the endocardium and by defining an improved initial region of interest which better includes the apex; Conduct a more elaborate evaluation of the proposed method by asking radiology technicians to manually segment exams for comparison with our segmentations and by comparing our blood volumes and ejection fractions with those obtained using ecocardiography, considered a gold standard for this kind of measures. Acknowledgments The authors would like to thank the radiology technicians at the Cardiology Service of Vila Nova de Gaia Hospital Center for their collaboration. The first author s work is funded by grant SFRH/BD/38073/2007 awarded by the portuguese Science and Technology Foundation (FCT). References 1. Fischbach, R., Juergens, K., Ozgun, M., Maintz, D., Grude, M., Seifarth, H., Heindel, W., Wichter, T.: Assessment of regional left ventricular function with multidetectorrow computed tomography versus magnetic resonance imaging. European Radiology 17(4), (2007) 2. Cury, R., Nieman, K., Shapiro, M., Nasir, K., Cury, R.C., Brady, T.: Comprehensive cardiac CT study: Evaluation of coronary arteries, left ventricular function, and myocardial perfusion is it possible? J. of Nuclear Cardiology 14(2), (2007) 3. Suri, J.S.: Computer vision, pattern recognition and image processing in left ventricle segmentation: The last 50 years. Patt. Analysis & App. 3(3), (2000) 4. Jolly, M.P.: Automatic segmentation of the left ventricle in cardiac MR and CT images. International Journal of Computer Vision 70(2), (2006) 5. Zheng, Y., Barbu, B., Gergescu, B., Scheuering, M., Comaniciu, D.: Four-chamber heart modeling and automatic segmentation for 3-D cardiac CT volumes using marginal space learning and steerable features. IEEE Transactions on Medical Imaging, Special Issue on Functional Imaging of the Heart 27(11), (2008) 6. Pannu, H.K., Flohr, T., Corl, F.M., Fishman, E.K.: Current concepts in multidetector row CT evaluation of the coronary arteries - pinciples, techniques, and anatomy. Radiographics 23, S111 S125 (2003)