École Nationale Supérieure de
Techniques Avancées

Abstract

Multi-class segmentation of prostate cancer (PCa) lesions in multi-parametric MRI if a difficult task that is mostly done using deep learning models trained with fully annotated dataset and standard loss function such as cross-entropy or dice. The aim of this work is first to train a network for the same task but using weak annotations and then to explore new loss functions taking into account the relationship between the different classes. After a study of the state of the art, we decided to use a method based on scribbles and constraints on the presence and the size of the objects to segment. We determined the optimal size and position of these scribbles, making our weakly supervised U-Net models perform similar to the fully supervised ones. It also allowed us to exploit a publicly available MRI dataset of PCa patient with weak annotations, resulting in the highest reported Cohen's kappa score for this task. Finally, we tried to use a custom kappa loss to increase the kappa on our fully supervised pipeline, but it is not successful yet.