Gastrointestinal endoscopic image classification using a hybrid modified inception network and a customized vision...

The Gastrointestinal (GI) tract plays a vital role in digestion by breaking down food into essential nutrients. Disorders such as bleeding lesions, ulcerative colitis, Inflammatory Bowel Disease (IBD), constipation, diarrhea, abdominal pain, nausea, and vomiting may indicate serious chronic conditions, including cancer. GI malignancies are among the leading causes of cancer-related mortality worldwide; however, early and accurate diagnosis can significantly reduce fatality rates. Existing endoscopic AI systems often struggle to generalize across datasets. They also face challenges in capturing both local lesion characteristics and long-range contextual information. To address this, the proposed study explores deep learning-based methods for automated classification of gastrointestinal diseases using endoscopic images. Two models were developed: a modified Inception architecture with four blocks and a customized Vision Transformer (ViT-3) consisting of three transformer encoder blocks. Deep features were extracted from both models, which were fused using a weighted fusion strategy. The redundant features were reduced using the Tree Growth Algorithm (TGA). The optimized feature set was then classified using machine learning classifiers to improve diagnostic performance. To evaluate its effectiveness, the framework was tested on two widely used benchmark datasets, Kvasir v1 and Kvasir v2, which include 4,000 and 8,000 images, respectively across eight GI disease categories. On the Kvasir v1 dataset, the proposed method achieved an accuracy of 98.9%, along with sensitivity, precision, and F1-score values of 98.87%, 98.86%, and 98.86%, respectively. Similar performance was observed on the Kvasir v2 dataset. These findings demonstrate that the proposed method provides a reliable and effective solution for early identification and classification of gastrointestinal diseases from endoscopic images.