Automated Colorectal Polyp Classification Using Deep Neural Networks with Colonoscopy Images

Colonoscopy screening test plays a crucial role in identifying and classifying possible cancerous polyps. The automation of polyp classification is challenging due to the limitations associated with the traditional handcrafted image features. This paper proposes a deep learning model for colorectal polyp classification, referred to as DeepCPC. This model consists of three main stages: image pre-processing using patch extraction and data augmentation; model initialization using six pre-trained CNNs to generate the fine-tuned baseline model; and formulating and learning generic yet discriminating image descriptors extracted and fused from the convolutional layers of two efficient CNNs architectures to classify colorectal polyps. The DeepCPC architecture has been fine-tuned on the CVC-Clinic dataset through a complete end-to-end training in which the patch extraction and image augmentation have been applied to generate more colonoscopy images for the patients. The experimental results show that the baseline fine-tuned model achieves an accuracy of 97.6%, and the final DeepCPC model achieves an accuracy of 98.4%. The reported results also demonstrate the capability of the proposed approach in identifying polyps in terms of precision, recall, and f-score. The DeepCPC helps the endoscopic physicians in classifying polyps and decreasing the colorectal polyp miss rate.