MNGLO-Net: A Cascade Network for Detecting Glomerulus and Classifying Pathological Types of Membranous Nephropathy

Membranous Nephropathy (MN) is a common adult nephrotic syndrome, whose key diagnostic process is Glomerular detection and classification. However, the development of MN diagnosis is impeded by the lack of pathologists and the complexity of glomerular pathological judgment. In this paper, we propose a two-stage cascade framework, MNGLO-Net, to identify three pathological types of glomeruli in our MNTG dataset. In first stage, NA-Net produces a detection of the glomerulus and classify the normal and abnormal types simultaneously. The resulting bounding box for each abnormal glomerulus is then used to crop a region of interest around the target in original image and the final feature maps of NA-Net to provide prior context to a refinement classification network, HL-Net, in second stage. Additionally, we propose the Color K-means guided Staining Transfer (CKST) module to solve the issue of single stain and differences among stains, and adopt the Prior knowledge guided Generating (PGAG) algorithm to deal with the challenge of significant shape difference. Experimental results show that MNGLO-Net reach the best results, i.e., 93.3% mean average precision (mAP) and 93.16% F1 score. Furthermore, we design the MN-GLOALY system to automatically produce pathological reports to promote the exploration of MN.