Multi-level information fusion for explainable diagnosis of melanoma using dermoscopic images

Melanoma is a prevalent and lethal form of cancer that is highly treatable when diagnosed at early stages. However, early detection remains challenging due to the subtle visual differences between malignant and benign skin lesions. We present a novel deep learning framework that integrates multi-level information fusion for explainable melanoma diagnosis. Our approach combines image segmentation and classification within a unified architecture, where segmentation masks highlighting clinically relevant regions are explicitly fused with classification features to guide diagnostic predictions. This fusion mechanism not only improves diagnostic performance but also provides clinically interpretable visual explanations that mimic the assessment process of expert dermatologists. Furthermore, we incorporate self-supervised learning to alleviate the dependency on large-scale annotated datasets, which are often costly and difficult to obtain in medical domains. Evaluated on the ISIC 2018 dataset, our framework achieves 83% AUC for melanoma classification compared to 76% for the baseline, while producing indicator localization masks that outperform existing methods across five clinical indicators. Notably, localization of rare indicators such as streaks improves from 5.16% to 18.66% Continuous Dice coefficient, and negative networks from 15.78% to 22.22%. The code is available at https://github.com/rusu4943/Bio-Unet.