MaxCAT-Net: Hybrid Transformer-Based Skin Lesion Classification and Lightweight Segmentation with Calibration and Cross-Dataset Validation

MaxCAT-Net (MaxViT-Tiny + CBAM + attention Transformer) achieves 0.869 accuracy (95% CI 0.847–0.890) on ISIC 2018 Task 3 test images (n = 1,405), exceeding several CNN and transformer baselines.

• Why it matters: Automated dermoscopic analysis is hindered by class imbalance, miscalibrated confidence, and limited robustness across imaging conditions and datasets.
• What they did: The model was trained on 10,012 de-duplicated images with a lesion-wise 72/14/14 split, then evaluated against multiple baselines; temperature scaling reduced ECE from 0.099 to 0.038.
• The result: Cross-dataset validation showed accuracy dropping to 0.341 on PAD-UFES-20 with melanoma recall 0.096, while HAM10000 melanoma recall was 0.587; segmentation Dice reached 0.903 (Swin-T U-Net) and 0.901 (FasterNet-T0 U-Net, ~1/10 parameters).

AI summary of the abstract · may contain errors · not medical advice · read the paper.