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

Abstract

Background: Early diagnosis of gastrointestinal (GI) cancer is crucial for patient prognosis, yet conventional methods suffer from invasiveness and insufficient molecular sensitivity. Raman spectroscopy offers non-invasive molecular fingerprinting, but spectral overlap in complex biological samples poses a challenge. This study introduces a diagnostic framework synergizing Raman spectroscopy with deep learning (specifically, convolutional neural networks -CNN) to quantitatively resolve spectral components for improved GI cancer detection.Results: Raman spectra from 829 GI tissues (760 benign, 69 malignant) and pure components (DNA, triolein, histone, collagen, actin) were collected. An improved CNN regression model, trained on 100,000 simulated spectra derived from the pure components, accurately quantified the relative proportions of these five biochemicals within tissue spectra (R² values: 0.91-0.98). Quantitative analysis revealed significantly higher coefficients for DNA, collagen, and actin, and lower coefficients for triolein and histone in malignant tissues compared to benign tissues (P < 0.01).Utilizing these quantitative molecular features, a LightGBM classification model achieved an accuracy of 97.2%, sensitivity of 90%, specificity of 98.1%, and an AUC of 0.973 on an independent test set of 579 samples.Significance: This work demonstrates a powerful approach for discriminating benign and malignant GI tissues by quantitatively modeling key molecular alterations using Raman spectroscopy and a tailored CNN. The high classification accuracy validates the clinical translational potential of this non-invasive method for GI cancer screening. Furthermore, the developed synergistic framework for quantitative spectral decomposition and classification offers a generalizable strategy extendable to other complex biological analyses, multimodal diagnostics, and potentially cancer staging.