Fast and Fully Automatic 3D Left Ventricular Segmentation Using Shape-Based B-Spline Explicit Active Surfaces

International audienceBackground, Motivation and ObjectiveCardiac function assessment is a critical step in cardiology and 3D ultrasound plays an increasingly important role. Automatic left ventricular (LV) segmentation remains however challenging particularly in the presence of artifacts and for images with low contrast-to-noise ratio (CNR). It is thus crucial to give segmentation tools prior information on the LVshape in order to fill in the gaps when image content is low. In this work, a fast automatic framework for full cycle LV segmentation is proposed that uses shape information from cardiac magnetic resonance (cMR) for a more accurate segmentation.Statement of Contribution/MethodsThe proposed approach couples B-spline explicit active surfaces (BEAS), an image information approach, with a statistical shape model (SSM) that gives prior information about the LV shape. 289 cMR datasets from a multi-center study, DOPPLER-CIP, were used to build an SSM describing the shape variations of the LV at end-diastole (ED) and endsystole (ES). This SSM is then used within BEAS, restricting the segmentation to the shape variability seen in the SSM. The framework begins with an automatic initialization of the LV at ED. BEAS is then used with the ED SSM regularization. The segmentation at ED is propagated throughout the heart cycle using a localized anatomical affine optical flow. At ES, the tracking result is refined using BEAS with the ES SSM regularization. The framework was tested on the CETUS challenge data, a multi-center multi-vendor dataset with manual contouring performed by three experts at ED and ES. The automaticsegmentation was evaluated with mean average distance (MAD), Hausdorff (HD) and Dice.Results/DiscussionThe proposed framework took in average 11s for each full cycle segmentation and showed excellent segmentation results, outperforming all state of the art methods presented in the CETUS challenge. Fig. 1 shows the best (left) and worst (right) results at ED (top) and ES (bottom) compared to manual contouring. The proposed framework achieved MAD, HD and Dice of 1.81±0.59mm, 6.31±1.69mm and 0.909±0.034 at ED and 1.98±0.66mm, 6.95±2.14mm and 0.875±0.046 at ES. ED and ES volumes and ejection fraction showed correlations of 0.953, 0.960 and 0.911 with the manual contours. Concluding, a fast and automatic LV segmentation framework based on BEAS and SSM was proposed,outperforming other state of the art LV segmentation tools