Use of subgraph mining in histopathology image classification

Breast cancer is the most common cancer in women and has a high mortality rate. Computer vision techniques can be used to help experts to analyze the breast cancer biopsy samples better. Graph neural networks (GNN) have been widely used to solve the classification of breast cancer images. Images in this field have varying sizes and GNNs can be applied to varying sized inputs. Graphs can store relations between the vertices of the graph and this is another reason why GNNs are preferred as a solution. We study the use of subgraph mining in classification of regions of interest (ROI) on breast histopathology images. We represent ROI samples with graphs by using patches sampled on nuclei-rich regions as the vertices of the graph. Both micro and macro level information are essential when classifying histopathology images. The patches are used to model micro-level information. We apply subgraph mining to the resulting graphs to identify frequently occurring subgraphs. Each subgraph is composed of a small number of patches and their relations, which can be used to represent higher level information. We also extract ROI-level features by applying a sliding window mechanism with larger sized patches. The ROI-level features, subgraph features and a third representation obtained from graph convolutional networks are fused to model macro-level information about the ROIs. We also study embedding the subgraphs in the graph representation as additional vertices. The proposed models are evaluated on a challenging breast pathology dataset that includes four diagnostic categories from the full spectrum. The experiments show that embedding the subgraphs in the graph representation improves the classification accuracy and the fused feature representation performs better than the individual representations in an ablation study.