Interpretable patient-voting deep learning-enhanced Raman spectroscopy of serum for breast Cancer detection

Early identification of breast cancer is essential for improving survival rates, yet current screening approaches often exhibit inadequate specificity or excessive invasiveness. Serum Raman spectroscopy (RS) provides a quick, non-destructive option, but its clinical use is impeded by the complexity of spectral data interpretation. Herein, we presented an interpretable deep learning framework, a one-dimensional convolutional neural network with a patient-voting strategy(PV-CNN), to evaluate serum RS from 732 individuals (366 patients and 366 healthy controls). Our model achieved a diagnostic accuracy of 95.21%, a sensitivity of 92.38%, and a specificity of 97.00%, significantly surpassing the performance of conventional ML algorithms. Furthermore, we employed Grad-CAM and SHAP analyses to elucidate the decision-making processes of deep learning, representing a significant advancement in addressing the "black-box" issue. This interpretability analysis found that tryptophan (1017 cm^-1) and phenylalanine (1002 cm^-1) were key Raman spectral indicators for breast cancer diagnosis. This study demonstrates that interpretable deep learning-enhanced RS can serve as a reliable, label-free, and physiologically explainable method for breast cancer detection.