Patient-Level Hyperspectral State-Space Learning for Melanoma Pathology Diagnosis

Histopathologic differentiation of cutaneous melanoma from pigmented nevus can be difficult when morphology is subtle. Microscopic hyperspectral imaging records tissue in contiguous narrow spectral bands across the visible and near-infrared range and may provide quantitative optical information beyond routine bright-field microscopy. We developed MelanoSpec-SSM, a patient-level spectral-spatial state-space algorithm for label-free hyperspectral pathology diagnosis.

This secondary methodological analysis used frozen sections from 100 patients (50 melanoma and 50 pigmented nevus) imaged from 400 to 1000 nm in 125 bands. After white-dark reflectance calibration, tissue extraction, spectral denoising, and strictly patient-level partitioning, MelanoSpec-SSM was compared with spectral classifiers, convolutional baselines, and a spectral-spatial transformer. The algorithm combines spectral state-space encoding, spatial patch learning, gated fusion, and attention-based multiple-instance aggregation.

In patient-level cross-validation, MelanoSpec-SSM achieved an accuracy of 0.940, sensitivity of 0.940, specificity of 0.940, F1-score of 0.940, and an area under the receiver operating characteristic curve of 0.972. In the held-out test set, 19 of 20 patients were correctly classified, with an area under the curve of 0.980. Model saliency and wavelength-removal analyses consistently assigned high task-specific importance to the 500-675 nm interval, while complementary near-infrared differences were also observed.

MelanoSpec-SSM provides an interpretable patient-level framework for visible-near-infrared hyperspectral differentiation of melanoma and pigmented nevus. The findings are preliminary because they arise from a single-center cohort of 100 patients and require external, multi-scanner, and prospective validation before clinical use.