This paper introduces a detection-based framework to segment glomeruli from digital scanning image of light microscopic slide of renal biopsy specimens. The proposed method aims to better use the precise localization ability of Faster R-CNN and powerful segmentation ability of U-Net. We use a detector to localize the glomeruli from whole slide image to make the segmentation only focus on the most relevant area of the image. We explored the effectiveness of the network depth on its localization and segmentation ability in glomerular classification, and then propose to use the classification network with enhanced ability of localization and segmentation to construct and initialize a segmentation network. We also propose a weakly supervised training strategy to train the segmentation network by taking advantage of the unique morphology of the glomerulus. Both strong initialization and weakly supervised training are used to resolve the problem of insufficient and inaccurate data annotations and enhance the adaptability of the segmentation network. Experimental results demonstrate that the proposed framework is effective and robust.
