An intelligent diagnostic algorithm for Raman spectroscopy of gastrointestinal cancer based on component modeling

Area of Science:

• Biomedical Engineering
• Spectroscopy
• Computational Biology

Background:

• Early gastrointestinal (GI) cancer diagnosis is vital but limited by invasive conventional methods lacking molecular sensitivity.
• Raman spectroscopy offers non-invasive molecular fingerprinting, but spectral overlap in biological samples is a challenge.

Purpose of the Study:

• To develop a diagnostic framework combining Raman spectroscopy and a convolutional neural network (CNN) for quantitative spectral component analysis.
• To improve the detection and differentiation of benign and malignant gastrointestinal tissues.

Main Methods:

• Trained an improved CNN regression model on 100,000 simulated spectra based on Raman spectra from 829 GI tissues and five pure biochemical components.
• Quantitatively analyzed the relative proportions of DNA, triolein, histone, collagen, and actin.
• Employed a LightGBM classification model using the quantitative molecular features for tissue discrimination.

Main Results:

• The CNN model accurately quantified biochemical proportions (R^2: 0.91-0.98).
• Malignant tissues showed significantly higher DNA, collagen, and actin, and lower triolein and histone coefficients (P < 0.01).
• The LightGBM classifier achieved 97.2% accuracy, 90% sensitivity, 98.1% specificity, and 0.973 AUC on an independent test set.

Conclusions:

• The developed framework quantitatively models key molecular alterations for effective discrimination of benign and malignant GI tissues.
• This non-invasive approach demonstrates significant clinical potential for GI cancer screening.
• The strategy is generalizable for other complex biological analyses and diagnostics.