CRHC-Net: A Cross-Representation Hybrid Clustering Framework for Tumor Identification in Histopathology Images

In histopathology images, tumor detection has been an issue because of high-dimensional heterogeneity in tissues and a huge variability caused by scanners, staining protocols and acquisition settings. To overcome these difficulties, CRHC-Net (Cross-representation hybrid clustering) is proposed, which is a hybrid framework of deep convolutional feature embedding along with artificial descriptors to provide strong tumor detection in diverse domains. This framework involves centroid based density and spectral clustering plans to enhance clustering properties used for cross-representation fusion to perform joint modelling of semantic pattern on local tissue texture. CRHC-Net is tested on two publicly available datasets of histopathology, TCGA and CAMELYON16, both intra-domain and cross-domain. The results shows that CRHC-Net can reach Dice Similarity Coefficient of 0.912 (whole-tumor segmentation) and for cross-dataset the similarity coefficient is high (0.895 DSC), which is better than the traditional clustering algorithm and deep learning baselines.