Improving Deep Learning Shape Consistency with a New Loss Function for Left Ventricle Segmentation in Cardiac MRI

Guaranteeing anatomical shape consistency in cardiac magnetic resonance imaging for left ventricle segmentation is a complex task due to its shape-changing during the cardiac cycle, the low contrast and resolution of images, the size change between apical and basal slices, the similarity with nearby organs, and the presence of cardiomyopathies that can deform the heart. Although producing segmentations very close to the ones produced by experts according to standard evaluation metrics, deep learning networks still often produce anatomically inconsistent segmentations. In this work, we propose a new shape-based loss function that favors shape consistency. The loss function uses shape information extracted from distance maps estimated by the network. We validate our approach with the ACDC and Sunnybrook public datasets by using standard metrics as well as a shape similarity metric. The results indicate that the proposed loss is able to improve shape similarity and demonstrate good generalization ability, while presenting competitive performance in the standard evaluation metrics.