Detection of gastrointestinal tract disorders using deep learning methods from colonoscopy images and videos

Abstract

Colorectal cancer (CRC) is the world's third most common cancer, with the second highest fatality rate. It is primarily the result of lower gastrointestinal tract (GI) disorders. The prevention of CRC mainly depends on the early detection and treatment of anomalies in the lower GI tract. Colonoscopy is the gold standard device used for diagnosing abnormalities in the lower GI tract as well as identifying anatomical landmarks and bowel preparation scales. However, it is time-consuming, tedious, and prone to error process, especially for those hospitals in low resource settings. Therefore, in this research, a real-time automated detection, classification, and localization of lower GI tract pre-colorectal cancerous abnormalities were done. The proposed system enables real-time detection, classification, and localization of common pathology, anatomical landmarks, and bowel preparation scale from colonoscopy images. To do the research, data was gathered both online (at hyper k-vasir dataset) and locally from the Yanet Internal Specialized Center and the Ethio-Tebib Hospital. Data augmentation techniques were applied to increase the training dataset. The pre-trained transfer learning SSD, YOLOv4, and YOLOv5 object detection model was used to develop the system with minimal fine-tuning of the hyper parameters and their performance was compared. The Yolo v5 model achieves good precision, recall, and mean average precision (mAP), 99.071%, 98.064% and 98.8%, respectively, on the testing data set. The developed artificial intelligence-based module would have the potential to assist gastroenterologists and general practitioners in decision-making. Even though the proposed work achieved the best performance, further improvement is required by increasing the size of the dataset to include other GI tract disease diagnoses.