Gastric histopathology image segmentation using a hierarchical conditional random field

For the Convolutional Neural Networks (CNNs) applied in the intelligentdiagnosis of gastric cancer, existing methods mostly focus on individualcharacteristics or network frameworks without a policy to depict the integralinformation. Mainly, Conditional Random Field (CRF), an efficient and stablealgorithm for analyzing images containing complicated contents, cancharacterize spatial relation in images. In this paper, a novel HierarchicalConditional Random Field (HCRF) based Gastric Histopathology Image Segmentation(GHIS) method is proposed, which can automatically localize abnormal (cancer)regions in gastric histopathology images obtained by an optical microscope toassist histopathologists in medical work. This HCRF model is built up withhigher order potentials, including pixel-level and patch-level potentials, andgraph-based post-processing is applied to further improve its segmentationperformance. Especially, a CNN is trained to build up the pixel-levelpotentials and another three CNNs are fine-tuned to build up the patch-levelpotentials for sufficient spatial segmentation information. In the experiment,a hematoxylin and eosin (H&E) stained gastric histopathological dataset with560 abnormal images are divided into training, validation and test sets with aratio of 1 : 1 : 2. Finally, segmentation accuracy, recall and specificity of78.91%, 65.59%, and 81.33% are achieved on the test set. Our HCRF modeldemonstrates high segmentation performance and shows its effectiveness andfuture potential in the GHIS field.