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

Area of Science:

• Biomedical Spectroscopy
• Computational Biology
• Oncology

Background:

• Early diagnosis of gastrointestinal (GI) cancer is crucial for improving patient outcomes.
• Noninvasive molecular quantification methods are needed for accurate cancer detection.
• Raman spectroscopy offers a potential approach for analyzing biochemical compositions in tissues.

Purpose of the Study:

• To develop and validate a synergistic framework integrating Raman spectroscopy and deep learning (CNN) for noninvasive GI cancer detection.
• To quantitatively analyze molecular alterations in GI tissues using spectral decomposition.
• To establish a robust diagnostic strategy for GI cancer based on molecular signatures.

Main Methods:

• Raman spectra were acquired from 927 GI tissues (82 malignant, 845 benign).
• Reference spectra of five biochemical components were theoretically calculated.
• A LightGBM classifier was trained on 10 features (5 coefficients + 5 ratio features) derived from spectral decomposition, using SMOTE for class imbalance.
• SHAP analysis and t-SNE were employed for feature importance assessment and visualization.

Main Results:

• The LightGBM model achieved high diagnostic performance: 98.2% accuracy, 99.4% sensitivity, 96.9% specificity, and 0.996 AUC.
• DNA and its ratio features were identified as the most significant indicators for distinguishing between benign and malignant tissues.
• Cross-validation confirmed the model's stability and generalizability, with a mean AUC of 0.996 ± 0.003.

Conclusions:

• The developed framework provides a robust and accurate strategy for GI cancer detection through quantitative molecular alteration analysis.
• This approach demonstrates excellent diagnostic performance and generalizability for complex biospectral applications.
• The integration of Raman spectroscopy and deep learning holds significant promise for noninvasive cancer diagnostics.