On the Inclusion of Topological Requirements in CNNs for Semantic Segmentation Applied to Radiotherapy
Résumé
The incorporation of prior knowledge into a medical segmentation task allows to compensate for the issue of weak boundary definition and to be more in line with anatomical reality even though the data do not explicitly show these characteristics. This motivation underlies the proposed contribution which aims to provide a unified variational framework involving topological requirements in the training of convolutional neural networks through the design of a suitable penalty in the loss function. More precisely, these topological constraints are implicitly enforced by viewing the segmentation assignment as a registration task between the considered image and its associated ground truth under incompressibility condition, making them homeomorphic. The application falls within the scope of organ-at-risk segmentation in CT (Computed Tomography) images, in the context of radiotherapy planning.