An interpretable deep learning framework for multi-scale diagnosis of gastrointestinal conditions across adult and pediatric populations

The accurate and interpretable diagnosis of gastrointestinal (GI) conditions remains a significant clinical challenge, particularly across diverse populations such as adults and children. In this study, we propose a domain adaptive explainable Multi-Scale Convolutional Neural Network integrated with squeeze-and-excitation mechanisms, developed to enhance diagnostic accuracy while ensuring transparency through interpretable visualizations. The architecture utilizes multi-scale feature extraction to capture spatial patterns at varying resolutions and incorporates channel-wise recalibration to underscore clinically salient features. To improve trust and clinical applicability, we integrated Gradient-weighted Class Activation Mapping (Grad-CAM) for a graphical illustration of model predictions. Extensive experiments were conducted on two benchmark datasets, WCE Curated Colon Disease (WCECCD) for adult endoscopy and CP-CHILD for pediatric colonoscopy. Our proposed model achieved a testing accuracy of 99.17% and an F1-score of 99.00% on the WCECCD dataset (Dataset A), and 97.29% accuracy with a 96.67% F1-score on CP-CHILD (Dataset B). It outperformed several state-of-the-art CNN-based deep learning models, including DenseNet121, MobileNetV2, and XceptionNet. Our proposed model demonstrated a competitive performance with VGG-19, where it shows a slightly higher accuracy on dataset A. By balancing performance, generalizability, and explainability, it indicates a strong possibility for the integration into real-world clinical workflows, specifically in resource-constrained settings where interpretability and computational efficiency are critical. • Proposed a multi-scale CNN for pediatric GI abnormality detection and localization. • Integrated squeeze-and-excitation to enhance channel-wise feature representation. • Applied Grad-CAM to improve model interpretability via visual explanations. • Achieved superior performance on two benchmark GI datasets. • Outperformed several state-of-the-art models in accuracy, precision, and generalization.