Multi-class classification of colorectal cancer tissues using pre-trained CNN models

Colorectal cancer (CRC) is a significant public health concern worldwide. Its early detection is critical since it dictates treatment options and significantly impacts survival time. A pathologist can make a histological diagnosis based on histological images obtained from a colonoscopy biopsy. The traditional visual assessment is time-consuming and highly unreliable because of the subjectivity of the evaluation. On the other hand, current approaches primarily rely on the use of diverse combinations of textual features and classifiers, as well as transfer learning, to classify distinct organizational types. However, the classification remains difficult since histological pictures comprise various tissue types and properties. In this work, we propose a deep learning technique based on pre-trained Convolutional Neural Networks (CNNs) for distinguishing eight classes of adenocarcinomas from healthy tissues. We explored multiple CNN architectures such as RESNET50, INCEPTIONV3, VGG 16, VGG19, RESNET152V2 and heuristically searched the best architecture for CRC detection by varying different optimizers. It was found that INCEPTIONV3 is the best performing model with 89.8% average accuracy for eight CRC classes.