Cross-Scale Guidance Integration Transformer for Instance Segmentation in Pathology Images

Abstract

Goal: To assess the degree of adenocarcinoma, pathologists need to manually review pathology images. To reduce their burdens and achieve good inter-observer as well as intra-observer reproducibility, instance segmentation methods can help pathologists quantify shapes of gland cells and provide an automatic solution for computer-assisted grading of adenocarcinoma. However, segmenting individual gland cells of different sizes remains a difficult challenge in computer aided diagnosis. Method: A novel cross-scale guidance integration transformer is proposed for gland cell instance segmentation. Our network contains a cross-scale guidance integration module to integrate multi-scale features learned from the pathology image. By using the integrated features from different field-of-views, the decoder with mask attention can better segment individual gland cells. Results: Compared with recent task-specific deep learning methods, our method can achieve state-of-the-art performance in two public gland cell datasets. Conclusions: By imposing cross-scale encoder information, our method can retrieve accurate gland cell segmentation to assist the pathologists for computer-assisted grading of adenocarcinoma.