Deep Raman Quantitative Profiling and Augmented Features for Biologically Interpretable GI Cancer Detection

Early diagnosis of gastrointestinal (GI) cancer is critical. Raman spectroscopy combined with deep learning offers a noninvasive molecular quantification approach. This study developed a synergistic framework integrating Raman spectroscopy and convolutional neural networks (CNN) for GI cancer detection through quantitative spectral decomposition. Raman spectra were collected from 927 GI tissues (82 malignant, 845 benign), and the reference spectra of five pure biochemical components (DNA, triolein, histone, collagen, and actin) were obtained through theoretical calculations. Component coefficients were extracted from the tissue spectra. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied to malignant coefficients. Five biologically interpretable ratio features were introduced alongside the original five coefficients, yielding a total of 10 features. A LightGBM classifier discriminated between benign and malignant tissues by utilizing these 10 features from the balanced data set (545 benign vs 545 SMOTE-malignant). SHAP analysis assessed feature importance, and t-SNE visualized feature distributions. The LightGBM model achieved superior performance: 98.2% accuracy, 99.4% sensitivity, 96.9% specificity, and 0.996 AUC. DNA and its ratio features were identified as the most important. Cross-validation confirmed model stability (mean AUC = 0.996 ± 0.003). This framework establishes a robust GI cancer detection strategy via quantitative molecular alteration analysis, demonstrating excellent diagnostic performance and generalizability for complex biospectral applications.