Histopathological Classification of Colorectal Polyps using Deep Learning

Abstract

Early diagnosis and classification of colorectal polyps are critical in reducing the morbidity and mortality rate of colorectal cancer (CRC). This paper proposes an automated method for histopathologically classifying colorectal polyps from 7000 µm H&E-stained images. First, a number of state-of-the-art deep learning models are developed and fine-tuned using transfer learning and ImageNet pre-trained weights. Subsequently, a baseline architecture is selected by comparing the trained models, and its performance is then optimized using data augmentation methods such as rotation, rescaling, mixup and cutout. Moreover, an extended variant of the adaptive moment estimation (Adam) optimizer called rectified Adam (Radam) and label smoothing are also used to boost the model performance. Based on the experimentation results using an open dataset, the proposed method achieved an accuracy of 90%, a precision of 90%, a recall of 89% and an F1-score of 0.91%.