Gastric histopathology image segmentation using a hierarchical conditional random field

The existing convolutional neural networks (CNNs) applied in the intelligent diagnosis of gastric cancer usually focus on individual characteristics and network framework without a policy to depict the integral information. Particularly, conditional random field (CRF), an effectual and stable algorithm for analysing the contents of complicated images, is able to characterize the spatial relation in images. In this paper, a novel hierarchical conditional random field (HCRF) model based gastric histopathology image segmentation (GHIS) method is proposed, which can automatically localize abnormal (cancer) regions in gastric histopathology images obtained by the optical microscope to assist histopathologists in medical work. The HCRF model is built up with higher order potentials, including pixel-level and patch-level potentials, and graph-based post-processing is applied to further improve our segmentation performance. We retrain a CNN to build up our pixel-level potentials and fine-tune another three CNNs to build up our patch-level potentials for abundant spatial segmentation information. In the experiment, segmentation accuracy, recall and specificity of 78.91%, 65.59%, and 81.33% are achieved on a hematoxylin and eosin (H&E) stained gastric histopathological dataset with 560 images. Our study demonstrates high segmentation performance achieved by the HCRF model and shows effectiveness and future potential of the proposed GHIS method.