Screening Pathological Abnormalities in Gastrointestinal Images Using Deep Ensemble Transfer Learning

Globally, gastrointestinal cancers including colorectal and esophageal cause a substantial amount of deaths every year. Early detection of these disorders requires an accurate, rapid, and automated diagnosis approach. Almost all prior deep learning-based gastrointestinal duct analysis research was restricted to polyp identification. However, esophagitis, ulcerative colitis, and numerous pathological findings of gastrointestinal organs must be analyzed together with polyps. This study focused on the detection of pathological findings in gastrointestinal images. A deep ensemble transfer learning approach was introduced to screen pathological findings. Initially, the eight EfficientNet members were used separately to train pathological findings. The ensemble networks were then constructed by incorporating a number of modified softmax averaging mathematical formulae. The ensemble’s efficiency and efficacy across individual networks were then verified quantitatively. The proposed ensemble network correctly predicted samples misclassified in the individual networks. Consequently, it has a 96.40% accuracy, 96.60% precision, and a recall of 96.40% for pathological findings. The proposed work surpassed nearly all recent comparable research in terms of accuracy and efficiency.