The abstract is read at the publisher; the summary is JClub's.
CapsNets (DR-CapsNet and BP-CapsNet) show superior adversarial robustness to ResNet-18/50 and MedViT across PneumoniaMNIST, BreastMNIST, NoduleMNIST3D, BloodMNIST, and MNIST.
- Why it matters: Medical diagnostic models must remain reliable under adversarial perturbations, yet robustness comparisons across alternative architectures in clinical-style image tasks remain limited.
- What they did: The study trained CNNs, a ViT, and two CapsNets on five datasets, then tested on adversarial examples from PGD, FGSM, AutoAttack, Carlini & Wagner, and black-box transfer attacks, plus PGD adversarial fine-tuning.
- The result: CapsNets maintained more consistent latent features and Grad-CAM attention maps after perturbations, and PGD fine-tuning improved robustness for all models, with MedViT gaining the most and narrowing the gap.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.