Explainable AI for colorectal lesion classification using deep learning models with attention mechanism

In the context of identifying colorectal lesions in colonoscopy images, previous AI models have predominantly focused on identifying abnormalities without providing comprehensible explanations for their predictions. To address this limitation, this paper introduces a novel framework that utilises two customised deep convolutional neural networks which we have developed, namely: ColoRecNet and Attention-BasedColoRecNet. These models aim to enhance human comprehension by providing interpretable explanations for their predictions. The framework also incorporates Grad-Cam, a visualisation technique, to highlight the distinctive features associated with each lesion class. The performance evaluation of the models on the testing set demonstrates that the Attention-BasedColoRecNet model surpasses the ColoRecNet model in terms of overall accuracy (95.67%), precision (96.02%), and recall (92.47%). Furthermore, the visual explanations generated by Grad-CAM heatmaps serve as additional validation, reinforcing that the Attention-BasedColoRecNet model possesses improved discriminative power and feature representations, resulting in superior classification performance.