MIDL Vision 2021

Morphology-based Losses for Weakly Supervised Segmentation of Mammograms Mickael Tardy is a PhD student at Ecole Centrale de Nantes under the supervision of Diana Mateus and a research engineer at French start-up Hera-MI. His paper explores the use of morphological operations to constrain segmentation output. He speaks to us ahead of his presentation today. This work proposes easy ways to perform weakly supervised segmentation with neural networks . By training a network to reproduce morphology operators , you can get rid of all the unnecessary noise and activations, creating nice shapes that often correspond to what you are looking for. Other papers have tried to create trainable kernels , but this one puts the morphology operations in the losses themselves. The method has been tested on mammography , which is high resolution and noisy imaging, so harder to segment. This is an easier way of optimizing the network and, as Mickael tells us, it yields quite surprising results. “ In the paper, we show that we got a gain of +14 points of DICE, so it wasn’t great at the very beginning, but it got considerably better after applying these losses, and it started to be much more meaningful , ” he explains. “ Without these losses, we get very small objects everywhere. Creating these losses allows us to have about four objects per image on average , which is quite cool. It’s probably not great for the production-ready solution but it’s already a very good step forward. ” 24 Presentation VISION MIDL