This study presents a machine learning framework for classifying skin lesions from Raman spectroscopy data, enabling non-invasive early-stage skin cancer detection. A dataset of 1,200 Raman spectra (600 benign, 600 malignant) was collected and pre-processed using baseline correction and vector normalisation. A hybrid 1D-CNN–BiLSTM architecture achieved 96.3 % accuracy, 95.8 % sensitivity, and 97.1 % specificity on stratified 5-fold cross-validation, outperforming conventional classifiers (SVM, Random Forest) by 2–4 accuracy points. Explainability was addressed through gradient-weighted class activation mapping (Grad-CAM), revealing spectrally meaningful regions of interest consistent with known biochemical markers of malignancy. The framework demonstrates clinical potential as a rapid, objective screening tool at the point of care.
Keywords: Raman spectroscopy · Skin cancer · CNN · BiLSTM · Deep learning · Biomedical optics