Deep Learning-Based Detection and Classification of Gastrointestinal Tract Diseases in Endoscopy Images

In the field of modern medicine, the detection and classification of gastrointestinal (GI) tract diseases from endoscopy images is a critical task since it helps in early intervention and improves patient outcomes. This study presents a unique way to deal with this topic, building on deep learning and using EfficientNetB1 and B0 architectures on the well-known Kvasir V2 dataset. The test performance of the EfficientNetB1 model is extremely satisfying, as it achieves an amazing accuracy of 94.75%, almost equal to the test accuracy of the EfficientNetB0 model, which is 94.31%. Although the accuracy of classification greatly varies from disease to disease, it demonstrates the complexity of GI tract conditions. Nevertheless, many finer distinctions were difficult to establish, such as stained-elevated polyps, dyed-resection margins, and esophagitis, besides typical variations including the normal appendix, normal pylorus, and normal z-line. These challenges necessitate the sophistication and expansion of our models by increasing the dataset to cover more pathological and normal cases. Yet, this study reveals the capability of deep learning to help physicians by providing them with advanced decision support devices for quick and precise GI disease diagnosis through endoscopy images. This study widens the horizons of ML-based diagnostic systems in gastroenterology, aiming at better therapy and prognosis for patients.