Optimization of a variational model using deep learning: An application to brain tumor segmentation

We propose a new variational model for saliency detection in images and its application to brain tumor segmentation. By incorporating a saliency term to a classical Total Variation based restoration functional, the new model is able to discriminate what is relevant (salient) from the background. The resulting non-convex and non-smooth problem has a special Difference of Convex structure and it is numerically solved using a proximal point algorithm that involves using a primal-dual method to solve an internal non-smooth sub-problem. We introduce a Deep Learning framework for using available knowledge from a specific application to optimize the parameters of the energy functional. The proposed framework is successfully tested on a database of Magnetic Resonance Images of the brain of patients suffering from a glio-blastoma reaching a Dice Similarity Coefficient of 85,7%.