Probability-Based Nuclei Detection and Critical-Region Guided Instance Segmentation

Nuclear instance segmentation in histopathological images is a key procedure in pathological diagnosis. In this regard, a typical class of solutions are deep learning-like nuclear instance segmentation methods, especially those based on the detection of nuclear critical regions. The existing instance segmentation methods based on nuclear critical regions are still insufficient in detecting and segmenting adhesion nuclei. In this study, we proposed a Critical-Region Guided Instance Segmentation (CGIS) method to precisely segment adhesion nuclei. Specifically, CGIS embed the critical region of an instance into the original image to guide the model to segment only the target instance, and provide an accurate segmentation mask for each nucleus. To improve the accuracy of critical region detection in CGIS, we proposed a boundary-insensitive feature, Central Probability Field (CPF). The CPF is calculated based on global morphological characteristics of nuclei, and can reduce the negative effects from the unreliable nuclear boundary details on the nuclear critical regions detection. We evaluated the effectiveness of the proposed CGIS and CPF feature on the Lizard and PanNuke datasets. It is shown that the proposed CGIS method can accurately segment adhesion nuclei, and the CPF feature can efficiently detect the critical regions of nuclei.