Improving histology images segmentation through spatial constraints and supervision

Abstract

We introduce two approaches to improve an existing color segmentation technique based on a Split and Merge quantization process for the study of stained histological images. We propose to modify the merge criterion : first, we include a spatial constraints heuristic; then we suggest the use of supervision and a more elaborated visual features representation. We tested these approaches on a renal biopsies dataset to automatically quantify interstitial fibrosis and show that supervision brings very significant improvements.