Oriented Cascade Mask R-CNN for Biomedical Image Segmentation

Biomedical images play a very important role in the diagnosis and treatment of diseases. In recent years, the technology of computational pathology has developed rapidly. It can effectively help doctors diagnose illnesses, which not only reduces the workload, but also achieves remarkable results. How to accurately segment cells in biomedical images is a very basic but also very important work. In order to analyze multi-tissue histology images, we propose a cell segmentation network, called Oriented Cascade Mask R-CNN (Region-CNN). It is a cascade network based on Oriented Mask R-CNN and Cascade Mask R-CNN. The model can simultaneously segment and classify cells in immunohistochemistry (IHC) images, and achieve good results. Oriented Cascade Mask R-CNN is a two-stage instance segmentation model. The whole network can be trained end-to-end and automatically segment cells. On our IHC dataset, the classification accuracy and recall are 85.8% and 86.0% respectively, and the mAP is 78.5%. Compared with Mask R-CNN, the mAP is improved by 4.9%, and the accuracy and recall are respectively improved by 9.5% and 10.2%.