Cell Detection by Robust Self-Trained Networks

Abstract

Cell nuclear detection on digital histopathology images plays an important role on computer-assisted cancer diagnostics. However, lack of manual annotations and variability of cells bring great challenges to fully-supervised learning. Therefore, we propose a Robust Self-Trained Network (RSTN) for cell detection. The backbone is an encoder-decoder trained by distance maps (DMs) generated from dot annotations of nuclei. To save manual efforts, RSTN is designed to involve reliable predicted DMs in optimization and detect cell centers for unknown images automatically. RSTN gains robustness by regularizing the network by dynamic graphs of DM patches. It exploits underlying graph structures and recognizes complex spatial patterns to locate cells of various shapes and colors. Experimental results show that it outperforms several classic and advanced models on both simulated fluorescence microscope images and real pathology slides for cell detection.