Combining Convolutional Neural Networks and Anatomical Shape-Based Priors for Cardiac Segmentation

We investigate whether leveraging high-resolution semantic segmentation from convolutional neural networks on Cardiac Tomography Angiography imaging, coupled with a shape-prior-based segmentation capable of enforcing the anatomical correctness can provide improved segmentation capabilities. While fully integrated approaches may be devised in principle, we investigate a simpler three-step approach for ease of implementation where, after leveraging a convolutional network to produce initial labels, we re-segment the labels using a fully geometric shaped-based algorithm followed by a post-processing refinement via active surfaces. Following the semantic segmentation, our second step is capable of generating a topologically correct cardiac model, albeit with lower resolution compared to the input labels, and is therefore capable of repairing any non-anatomical mislabeling. The post-processing step then recaptures the lost small-scale structure making the combined strategy successful in recovering a topologically correct segmentation of the imaging data of quality comparable, if not superior, to the initial labels. Our results show dice scores comparable to those obtained by using deep learning alone but with much improved performance in terms of Hausdorff distance due to the removal of erroneous islands and holes which often evade notice using only dice scores. In addition, by design, our segmentation is topologically correct. This preliminary investigation fully demonstrates the advantages of a hybrid semantic-geometric approach and motivates us in pursuing the investigation of a more integrated strategy in which semantic labels and geometric priors will be integrated as competing penalty terms within the optimization algorithm.