GI-Vision XAI: Duplicate-Grouped Evaluation of a Four-Class ResNet-18 Endoscopy Image Classifier with Grad-CAM and a Desktop Research Interface

GI-Vision XAI is a four-class gastrointestinal endoscopy image classification research prototype combining a ResNet-18 convolutional neural network, duplicate-grouped image partitioning, Grad-CAM visualizations, and a PySide6 desktop interface. The manuscript reports archived test-set results for 1,805 images: 97.06% accuracy, 96.98% macro F1-score, 96.98% macro sensitivity, 99.03% macro specificity, and 99.87% macro one-vs-rest ROC AUC. These results are reported from archived aggregate artifacts and were not independently reproduced from the original images and prediction-level outputs. The evaluation is image-level and does not establish patient-level independence, external generalization, calibrated confidence, validated lesion localization, or clinical diagnostic validity. The software is a research prototype, not a clinically validated diagnostic system. The original hospital-sourced images are restricted and are not included in this deposit. Any access to the source data remains subject to applicable permissions, institutional policies, and privacy and ethics requirements.