Multiscale transformer model based on Raman spectroscopy detection in diagnosis of colorectal cancer

There is currently a lack of low-cost, nonintrusive methods for colorectal adenocarcinoma (CRA) detection. Moreover, the application of Raman spectroscopy (RS)-based CRA detection is limited due to its high noise levels and low characterization abilities.

Propose a method for the early screening of CRA based on RS and deep learning techniques.

An MS-Former (multiscale transformer) method was constructed by combining multiscale convolution and self-attention mechanisms to adapt spectral scenarios based on RS and surface-enhanced Raman spectra (SERS) at the cellular and patient serum levels, respectively. Furthermore, Raman spectral critical frequency test (classifying the spectrum extracted from each threshold by linear discriminant analysis (LDA) using the classical machine learning method) was performed and the concerning Raman shifts were screened in the identification process of different samples.

At the cellular level, the proposed model can identify normal cells and cancer cells at an accuracy of 99.12%. At the serum level, for the first time, this model achieved AUC values of 98.45% and 99.40% in identification of adenoma and adenocarcinoma, respectively. When classifying patients into groups of healthy, adenoma, and adenocarcinoma, the single Raman detection and individual level accuracy were 94.78%, 100%, respectively. Matching metabolites were obtained through matching the screened Raman shifts, which greatly improved the interpretability and biological significance of the results.

The proposed model can conveniently and rapidly screens patients with colorectal adenomas and adenocarcinomas, requiring no pretreatment and exhibiting extremely high accuracy and low cost.