RESNEXT-BASED DEEP LEARNING FRAMEWORK FOR AUTOMATED GASTROINTESTINAL ENDOSCOPY IMAGE CLASSIFICATION WITH RU-NET SEGMENTATION

Abstract

In order to improve patient outcomes and lower diagnostic mistakes during endoscopic examinations, gastrointestinal (GI) illnesses must be detected early and accurately. This study presents an intelligent deep learning framework for automated analysis of GI endoscopy images by combining image preprocessing, G-Net Light feature extraction, RU-Net-based lesion segmentation, and advanced convolutional neural network classifiers. A comparative evaluation of ResNet and ResNeXt was performed to assess their effectiveness in lesion classification. The experimental results indicate that ResNeXt provides superior performance by capturing richer multi-scale features through grouped convolutions while maintaining computational efficiency. In comparison to ResNet, which obtained 96.1% accuracy, 95.8% precision, 95.4% recall, 95.6% F1-score, and an AUC of 0.975, ResNeXt produced a classification accuracy of 97.8%, precision of 97.4%, recall of 97.1%, F1-score of 97.2%, and an AUC of 0.989. The suggested segmentation framework also showed good localization performance, with an Intersection-over-Union (IoU) of 90.2% and a Dice Similarity Coefficient of 94.6%Furthermore, ResNeXt reduced the classification error by approximately 42% compared with ResNet while requiring only a marginal increase in computational time. These results highlight the effectiveness of integrating RU-Net segmentation with ResNeXt classification for reliable, accurate, and clinically applicable gastrointestinal lesion detection and decision support.