Tumor-Stroma Classification in Colorectal Cancer Patients with Transfer Learning based Binary Classifier

Computer aided diagnosis (CAD) of the colorectal diseases based on histological images of the cancer tissues have attracted massive interests in digital pathology. In this context, we propose, in this paper, a transfer learning approach for solving binary classification task of the tumor-stroma region of colorectal cancer (CRC) patients. The pretrained convolutional neural networks (CNN) were used with ImageNet dataset as a source dataset, where the features of the bottleneck layer are transferred to the simple classification head. The classification head consists of a global maxpool layer followed by dropout and a single unit dense layer that outputs raw prediction values which are ultimately classified as 0 or 1, each denoting tumor and stroma region respectively. A target dataset containing small number of CRC histological images is used to train the classification head. During inference, the proposed binary classifier accurately classifies the tumor-stroma region and significantly improves the classification accuracy compared to reference schemes.