G-NET LIGHT AND RU-NET LIGHT-BASED DEEP LEARNING FRAMEWORK FOR GASTROINTESTINAL ENDOSCOPIC IMAGE ANALYSIS

Abstract

Gastrointestinal endoscopy plays an important role in the early detection of inflammatory abnormalities and lesions; however, manual interpretation of endoscopic images is often affected by visual complexity and inter observer variability. This study presents a deep learning-based framework for automated gastrointestinal endoscopic image analysis by integrating lightweight and residual segmentation networks with convolutional neural network-based classification and explainable artificial intelligence. The segmentation performance of G Net Light and Residual U-Net (RU-Net Light) was comparatively evaluated using Dice coefficient, Intersection over Union (IoU), and pixel accuracy. In the overall segmentation analysis, G-Net Light achieved a Dice coefficient of 0.234, IoU of 0.146, and pixel accuracy of 84.7%, whereas RU-Net Light substantially improved these values to 0.726, 0.608, and 91.8%, respectively. For oesophagitis segmentation, RU-Net Light further obtained a Dice coefficient of 0.756, IoU of 0.637, and pixel accuracy of 92.5%, compared with 0.426, 0.298, and 86.7% achieved by G-Net Light. These findings demonstrate the effectiveness of residual learning and encoder decoder skip connections in preserving lesion boundaries and extracting fine-grained mucosal features. Among the evaluated classification models, Xception achieved the highest accuracy of 93.33%, precision of 93.46%, recall of 93.09%, and F1-score of 93.14%. The integration of RU-Net Light, Xception, and Grad-CAM provides an accurate and interpretable framework for automated gastrointestinal lesion analysis and has potential to support gastroenterologists in reliable computer-assisted diagnosis.