Multi-loss convolutional networks for gland analysis in microscopy

Abstract

Manual tissue diagnosis is the most prevalent approach to cancer diagnosis. However, it mainly relies on a subjective visual quantification of specific morphometric features, which often leads to a relatively limited reproducibility among experts. In most computational techniques proposed to automate the diagnostic procedure, accurate segmentation is paramount as a precursor to the extraction of relevant morphometric features. Since the ultimate goal of segmentation is generally classification, yet a given class imparts an expected tissue appearance beneficial to segmentation, we pose the problem of automatic tissue analysis as the joint task of segmentation and classification. We propose a novel multi-objective learning method that optimizes a single unified deep fully convolutional neural network with two distinct loss functions. We illustrate our reasoning on the task of colon adenocarcinomas diagnosis and show how glands' classification can facilitate their segmentation by adding class-specific spatial priors. The final classification also benefits from this joint learning framework yielding an improvement of 6% over classification-only models.