A Hybrid 1D-CNN and Transformer Architecture for Differentiating Malignant Melanoma from Non-Melanoma Skin Cancers using Raman Spectroscopy

Abstract

Clinical differentiation between malignant melanoma (MM) and non-melanoma skin cancers (NMSCs) is critical for patient prognosis. Raman spectroscopy offers a noninvasive, molecule-sensitive approach, yet its complex spectral data requires sophisticated analytical tools. This paper proposes a novel hybrid deep learning architecture, combining a five-block 1D Convolutional Neural Network (1D-CNN) with a six-layer Transformer encoder, to leverage both local spectral features and their global contextual relationships. The 1D-CNN acts as a powerful local feature extractor, while the Transformer models long-range dependencies across the entire spectral range. Evaluated on a large clinical dataset of 1200 spectra (150 MM, 350 BCC, 350 SCC, 350 benign), our model demonstrates superior performance, achieving an overall accuracy of 96.8% and an AUC of 0.987. Notably, the sensitivity for detecting the most aggressive MM class reached 94.6%. These results significantly outperform baseline models, including a standalone 1D-CNN, highlighting the efficacy of the proposed hybrid architecture for robust, automated skin cancer classification.