Cardiac Left Ventricle Segmentation in CT and MR Using a Fuzzy Inferenced 3D Active Shape Model

Abstract 116 Purpose – Nowadays, Multi-Slice CT and MR are becoming the preferred methods to evaluate diseases affecting the heart. Both modalities are widely utilized for assessment of global cardiac function. Manual quantitative analysis of the vast amounts of image data is arduous. Consequently, automatic segmentation methods that need little-to-no human interaction are required. Material and Methods – A 3D Active Shape Model for automatic segmentation of cardiac left ventricle CT and MR volumes, without retraining its statistical component for each modality is presented. A fuzzy inference system was incorporated into the model to increase the robustness of segmentation of multiple modality-based data sets. In a two-step approach, a volume of interest is first classified into three tissue classes (blood, myocardium, air). Secondly, tissue transitions are detected and shape constraints are applied to the detected volume to affirm, based on the statistical framework, that a plausible human heart shape is obtained. Results – Evaluation was done on 15 MR and 23 CT data sets. Pairs of automatic and expert segmentations were compared in 100 locations per slice. Results of point-to-curve distance measurements between automatic and manual surfaces are presented in Table 1. For MR, 93.3% of epicardial contours and 91.5% of endocardial contours had a maximum error of 5 mm along 95% of the contour arc length. For CT, those numbers were 82.4% (epicardium) and 74.1% (endocardium). Volume regression analysis revealed good linear correlations between manual and automatic volumes, r2 ‡ 0.98. Conclusion – This study shows that our 3D-ASM is a robust promising instrument for automatic cardiac left ventricle segmentation. Without retraining its statistical component, it is applicable to routinely acquired CT and MR studies.