Hyperspectral imaging-based framework for automated detection of dysplastic melanocytes in dermatopathology

Hyperspectral imaging provides rich spectral information that can be leveraged for non-invasive analysis of skin lesions. This study proposes a framework to automatically detect melanocytes in hyperspectral skin images and identify the most informative spectral bands for lesion classification. A total of 15 hyperspectral images representing three nevus, dysplastic nevus, and melanoma were analyzed. Melanocytes were segmented using unsupervised clustering, applied both to the original hyperspectral data and the a* chrominance. Following post-processing, individual melanocyte spectra were extracted in the 760 – 1000 nm near-infrared range. These spectra were used to train and evaluate four supervised classifiers, along with two stacked ensemble models, developed to boost performance. The best-performing classification experiment, which achieved an accuracy of 0.73, served as the basis for identifying relevant near-infrared wavelengths for skin lesion discrimination. Results demonstrate that this pipeline enables both melanocyte detection and spectral feature selection, supporting its potential for diagnostic decision support in dermatology.