Automated colorectal polyp detection based on image enhancement and dual-path CNN architecture

Abstract

Colorectal Cancer (CRC) has the highest mortality rate of all cancers and is currently the third leading cause of cancer-related death worldwide. The early detection and diagnosis of colorectal polyps are necessary for early interventional therapies. The use of AI and ML techniques to analyse colonoscopy images has been gaining traction in recent years for early and accurate detection of polyps and other colorectal abnormalities. Existing deep learning classification and detection methods of polyps are computationally intensive, restrict memory potency, require extensive training, and affect the optimization of hyperparameters. This makes them unsuitable for real-time applications and applications with limited computing resources. This paper proposes a Dual-Path Convolutional Neural Network (DP-CNN) to classify polyp and non-polyp patches from the colonoscopy images. The proposed approach comprises image enhancement followed by the use of DP-CNN architecture and a sigmoid classifier for efficient detection of polyps. The publicly available database CVC ClinicDB is used to train the proposed network, and it is tested on ETIS-Larib and CVC ColonDB databases. The testing accuracy of the network on CVC ColonDB and ETIS-Larib are 99.60%, 90.81%, respectively. The performance measures are as follows: precision (100%), recall (99.20%), F1 score (99.60%) and F2 score (99.83%) on CVC ColonDB database and precision (89.81%), recall (92.85%), F1 score (91.00%) and F2 score (89.91%) on ETIS-Larib database. Compared with other existing methods, the proposed approach outperforms in precision, recall, F1-score, and F2-score in both databases. The number of learnable parameters of the proposed method is 8737. The proposed approach is promising as an accurate polyp detection technique. It is applicable for real-time applications due to lower complexity and fewer learnable parameters than required by other existing methods.