Focal Dice Loss-Based V-Net for Liver Segments Classification

Liver segmentation is a crucial step in surgical planning from computed tomography scans. The possibility to obtain precise delineation of the liver boundaries with exploitation automatic techniques can help radiologists, reducing annotation time and providing more objective repeatable results. Subsequent phases typically involve vessels’ segments’ classification. It especially important recognize different segments, since each has its own vascularization, so, hepatic segmentectomies be performed during surgery, avoiding unnecessary removal healthy parenchyma. In this work, we focused on classification task. We exploited 2.5D Convolutional Neural Network (CNN), namely V-Net, trained multi-class focal Dice loss. idea loss was originally thought as cross-entropy function, aiming at focusing “hard” samples, gradient being overwhelmed by large number falsenegatives. paper, introduce two novel formulations, one based concept individual voxel’s probability another related formulation for sets. By applying aforementioned task, were able respectable results, an average coefficient among classes 82.91%. Moreover, knowledge anatomic configurations allowed application set rules post-processing phase, slightly improving final obtaining 83.38%. accuracy close 99%. best model turned out conducted Wilcoxon signed-rank test check if these results statistically significant, confirming their relevance.