Explainable Weakly Supervised Deep Learning Model for Anomaly Classification and Localization in Wireless Capsule Endoscopy Images

ABSTRACT Clinically reliable AI models are an increasingly important development in Wireless Capsule Endoscopy (WCE) image analysis due to the large volume of labeled samples. Although various computer vision architectures have been proposed for WCE image classification and anomaly localization, challenges remain in achieving domain‐specific lesion localization and enhancing the interpretability of model predictions. To overcome these issues, we present a weakly supervised deep learning framework that integrates self‐attention‐augmented convolutional neural networks (CNNs) to improve both classification accuracy and visual explainability. Our approach uses attention‐augmented convolutional features to discriminate lesion regions from surrounding tissue. A lesion‐specific saliency map is generated by combining attention responses with Gradient‐weighted Class Activation Mapping (GradCAM), enabling effective and interpretable localization of anomalies in a weakly supervised manner. The perturbation‐based analysis of these saliency regions facilitates a clear understanding of model decision‐making. The proposed classification framework achieves a classification accuracy of 95% and a localization accuracy of 85%, outperforming conventional CNN models, while the integration of the self‐attention mechanism increases the parameter count to 88 M and 12 GFLOPS. The proposed framework maintains a favorable trade‐off between complexity and accuracy. Results highlight the efficiency of attention mechanisms in improving lesion‐aware representation and visual explainability, contributing to the improvement of trustworthy AI in gastrointestinal diagnostics.