CBAM-enhanced MaxViT-Tiny for gastrointestinal endoscopy image classification: Robustness, explainability, and external validation

Automated classification of the huge number of images generated during gastrointestinal (GI) endoscopy is required for reliable and interpretable diagnosis, but there is concern about the vulnerability of deep models to adversarial perturbations. In the current study, a MaxViT-Tiny backbone with a Convolutional Block Attention Module (CBAM) head is used for 8-class classification of the Kvasir endoscopic images. After deleting duplicate images, 3,996 images remained, which were split into a training set (70%), a validation set (15%) and a test set (15%); the model was compared with five backbones using the same CBAM head under three random seeds. The proposed model obtained the highest test accuracy of 94.33%, a macro F1 score of 0.9431, and a macro area under the receiver operating characteristic curve (AUC) of 0.9975, with a mean accuracy of 93.72 {+/-} 1.06% across all seeds. White-box FGSM, PGD, CW-L2, and MI-FGSM attacks greatly reduced the accuracy: after adversarial fine-tuning, the accuracy under FGSM was 74.17%, while the accuracy under PGD and MI-FGSM was below 4%, showing poor cross-attack generalization. A preliminary attack detector obtained an AUC of 0.857. Grad-CAM++, Score-CAM, and GradientSHAP sometimes identified clinically relevant regions, but also sometimes focused on non-anatomical artefacts. High accuracy on clean images, even with attention augmentation, therefore does not imply high accuracy on adversarial images or reliable explanations, and the model needs further verification on independent datasets and multi-centre validation before clinical use.