Work

MelanoFusion: An Explainable and Uncertainty-Aware Hybrid CNN–Transformer Framework for Dermoscopic Melanoma Detection

Year: 2026
Type: preprint
Abstract: Melanoma is an aggressive skin cancer requiring early detection, yet automated analysis remains affected by lesion variation, calibration errors, uncertainty, and subgroup disparities. This study presents MelanoFusion, a hybrid framework integrating explainability, uncertainty estimation, and fairness-oriented representation learning for melanoma detection. It combines EfficientNetV2-S and Swin Transformer encoders with cross-attention, lesion attention, feature-pyramid processing, multi-scale transformer layers, and an uncertainty head. Images were resized to 224 × 224 and processed using hair removal, median filtering, contrast enhancement, colour normalization, and watershed-based lesion priors; U-Net was used for lesion segmentation. MelanoFusion was evaluated against DenseNet121, ResNet101, EfficientNetV2-S, ConvNeXt-Tiny, and Swin Transformer on HAM10000, ISIC 2018, and PH². Segmentation Dice was 0.878 on HAM10000 and 0.830 on PH². MelanoFusion achieved ROC-AUC values of 0.687, 0.740, and 0.844 on HAM10000, ISIC 2018, and PH², respectively. Calibration was favourable on HAM10000, with an expected calibration error of 0.116. Uncertainty was significantly associated with classification error, while subgroup analysis showed significant sensitivity differences across age groups but not sex. Ablation results indicated that several proposed components did not improve discrimination. The findings support joint evaluation of discrimination, calibration, uncertainty, explainability, and subgroup behaviour, while indicating the need for external validation and clinical evaluation before deployment.
Language: English
Subfield: Oncology
Field: Medicine
Open Access status: green
OpenAlex ID: W7236005006
Cites: 30
Cited by:
Related to: 10