Biophysically interpretable basal cell carcinoma classification using Raman spectroscopy transformer model

Raman spectroscopy is a non-invasive technique for imaging skin tissues and has shown high accuracy in discriminating between cancerous and noncancerous skin. Analysis of Raman spectra has typically used biophysical or machine learning (ML) models. While biophysical models provide interpretation, they can be difficult to develop and generalize across instruments and experimental settings. ML models can be simpler to implement and transfer across systems, but lack interpretability, especially with respect to spectral wave number regions that contribute most towards classification. This study aims to introduce an interpretable ML approach for Raman analysis through a transformer-based ML model for the classification of basal cell carcinoma (BCC). We employ tokenization to partition the spectral data, enabling the model to capture weighted biochemical features for classification. Our proposed strategy achieves a sensitivity of 92% and specificity of 96%, with a theoretical specificity of 67% at perfect sensitivity. In addition, our method allows interpretability through visualization of wave number regions that contribute most towards classification. We find bands at ∼856-940 cm⁻¹, ∼1128 cm⁻¹, and ∼1440 cm⁻¹ corresponding to keratin, collagen, elastin, and lipid-related structures to be attended to during classification. This approach aims to bridge the gap between ML predictions and underlying biophysical mechanisms, and holds promise for broader applications in optical diagnostics.