Ensemble learning for melanoma detection in hyperspectral skin images using chromophore-based features
Artem A. Borovkov1, Irina A. Matveeva1; 1Samara National Research University, Samara, Russia
Abstract
Melanoma remains the leading cause of skin cancer mortality, while the sensitivity of primary visual examination does not exceed 50 %. We propose a method for automatic identification of skin neoplasms from hyperspectral images that combines biophysical interpretation of spectra with deep learning. Data were acquired using an acousto-optic imaging spectrometer (151 spectral channels spanning 450–750 nm, 2 nm step) from 219 patients across three classes (melanoma, nevus, keratosis). The lesion is delineated by a U-Net taking the full hypercube as input; the median Dice coefficient reached 0.836. Within the mask, spectra are decomposed by non-negative matrix factorization into four components. The recovered basis, obtained without prior constraints, matched the absorption spectra of real skin chromophores (oxyhemoglobin with its 575 nm peak, deoxyhemoglobin, melanin) and scattering, confirming the physical validity of the model. Seventy-two statistical, textural (GLCM) and spatial features describing pigmentation heterogeneity are computed from the concentration maps. In parallel, an RGB projection of the hypercube is processed by the PanDerm dermatological transformer fine-tuned with a class-balanced contrastive loss. Both branches are fused by soft voting. On an independent test set of 66 patients, the proposed ensemble achieved an AUC of 0.910, outperforming the spectral feature-based model (0.874), the neural model (0.798), and classical spectral statistics (0.634). In screening mode (threshold 0.192) all 13 melanomas were detected at a specificity of 0.604. A key advantage of the proposed approach is interpretability, as predictions are linked to concentrations of physiologically meaningful chromophores rather than abstract features.
Speaker
Borovkov Artem
Department of Laser and Biotechnical Systems, Samara National Research University, 34 Moskovskoye Shosse, Samara 443086, Russia
Russia
Discussion
Ask question
