Nuclei classification using dual view CNNs with multi-crop module in histology images

Histopathology image diagnostic technique is a quite common requirement; however, cell nuclei classification is still one of key challenge due to complex tissue structure and diversity of nuclear morphology. Cell nuclei categories are often defined by contextual information, including central nucleus and surrounding background. In this paper, we propose a Dual-View Convolutional Neural Networks (DV-CNNs) that captures contextual contents from different views. The DV-CNNs are composed of two independent pathways, one for global region and another for center local region. Noted that each pathway with “multi-crop module” can extract five different feature regions. Common networks do not fully utilize the local information, but the designed cropping module catches information for more complete features. In experiments, two pipelines are complementary to each other in score fusion. To verify the performance in proposed framework, it is evaluated on a colorectal adenocarcinoma image database with more than 20,000 nuclei. Compared with existing methods, our proposed DV-CNNs with multi-crop module demonstrate better performance.