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

Area of Science:

• Biomedical Engineering
• Medical Diagnostics
• Computational Biology

Background:

• Early breast cancer detection is crucial for survival but faces challenges with current screening methods' specificity and invasiveness.
• Serum Raman spectroscopy (RS) offers a non-destructive diagnostic potential but requires advanced data interpretation techniques.
• The

Purpose of the Study:

• To develop and validate an interpretable deep learning framework for breast cancer diagnosis using serum RS.
• To address the data interpretation complexity and

Main Methods:

• A one-dimensional convolutional neural network with a patient-voting strategy (PV-CNN) was developed.
• Serum RS data from 732 individuals (366 patients, 366 controls) were analyzed.
• Interpretability techniques (Grad-CAM, SHAP) were employed to understand model decisions.

Main Results:

• The PV-CNN model achieved high diagnostic performance: 95.21% accuracy, 92.38% sensitivity, and 97.00% specificity.
• The model significantly outperformed conventional machine learning algorithms.
• Tryptophan (1017 cm⁻¹) and phenylalanine (1002 cm⁻¹) were identified as key spectral biomarkers.

Conclusions:

• Interpretable deep learning enhances serum RS for reliable, label-free breast cancer detection.
• The framework provides physiological explanations for diagnostic decisions, overcoming the "black-box" problem.
• This approach represents a promising advancement for non-invasive cancer diagnostics.