Semi-automatic Segmentation of Tissue Regions in Digital Histopathological Image

Segmentation of tissue regions in the digital histopathological images refers to the identification and segmentation of tissues such as epithelium, glandular cavity, fibers, etc. Precise segmentation of tissues is key to pre-determining the regions with the greatest diagnostic value, which can support clinical diagnosis, particularly with regard to etiology and severity. In view of the uneven quality of histopathological images and the difficulty of manual segmentation. In this paper, an approach based on weakly supervised learning and deep learning has been proposed to build a semi-automatic segmentation model of tissue regions. The model uses superpixel classification to pre-segment the tissues, the tissue region boundary is preserved, and the automatic segmentation of the tissues is finally achieved based on the deep convolutional neural network. The effectiveness of the model is evaluated on 600 cervical histopathology images provided by the hospital. The results show that the proposed method achieves 82.52% mean IoU of epithelial segmentation and 81.67% mean IoU of glandular lumen segmentation in cervical histopathological images. The model is superior to traditional manual feature representation methods and classical deep convolution neural network methods in segmentation accuracy and efficiency.