MelanoSpectraNet: A Deep Learning Framework for Multi-Stage Melanoma Detection Using Hyperspectral Imaging

<title>Abstract</title> <p>Melanoma is the most aggressive type of skin cancer, but it is very difficult to diagnose in the early stages because of its similarity to benign skin lesions and its ability to metastasize quickly. This paper presents a non-invasive melanoma staging system that combines Hyperspectral Imaging with an attention-based residual 3D U-Net model, allowing for the accurate classification of melanoma stages (I-IV) based on volumetric spectral-spatial data. Unlike traditional 3D U-Net models, which are based solely on conventional convolutional encoding-decoding paradigms and consider all features of the skip connections to be of equal importance, the proposed approach incorporates residual double convolution layers and attention-driven skip connections to improve feature representation. Residual learning facilitates gradient backpropagation and training of deep volumetric models, and attention gates help to automatically focus on the melanoma-related areas of interest and attenuate background noise in hyperspectral data. The experimental results show that the proposed model has a substantial improvement over the standard 3D U-Net baseline, achieving an overall accuracy of 97.94% and a weighted F1-score of 97.15%, especially for the differentiation between early and late-stage melanoma. The proposed model provides a promising starting point for the development of real-time, cost-effective, and clinically feasible melanoma staging systems. Future research will investigate more advanced spectral attention mechanisms and larger-scale datasets to further improve the diagnostic capabilities and individualized treatment strategies.</p>