MelanoSpectraNet deep learning framework for multistage melanoma detection using hyperspectral imaging

Abstract Melanoma is the most aggressive form of skin cancer, but due to its rapid metastasis and resemblance to benign skin lesions, it is extremely challenging to diagnose in its early stages. In order to accurately classify melanoma stages (I–IV) based on volumetric spectral-spatial data, this paper presents a non-invasive melanoma staging system that integrates Hyperspectral Imaging with an attention-based residual 3D U-Net model. The suggested method uses residual double convolution layers and attention-driven skip connections to enhance feature representation, in contrast to conventional 3D U-Net models, which are exclusively based on traditional convolutional encoding–decoding paradigms and treat all features of the skip connections as equally important. Attention gates assist in automatically focusing on the melanoma-related regions of interest and reducing background noise in hyperspectral data, while residual learning makes gradient backpropagation and deep volumetric model training easier. The experimental results imply that, based on hyper-spectral imaging data, the MelanoSpectraNet framework for the classification of melanoma in a phased way has been successful. The Attention-augmented 3D U-Net generated an accuracy of 97.94%, precision of 97.86%, recall of 97.48%, and F1-score of 97.15%. These findings demonstrate how well the MelanoSpectraNet architecture captures spectral spatial characteristics, enabling precise melanoma disease stage categorization.