AI-assisted Diagnosis of Early-stage Esophageal Cancer Probed by Raman Spectroscopy

Raman spectroscopy with chemometric and AI analysis is a promising technique for diagnosis of early stage cancerous biochemical changes. We employed Raman spectroscopy to probe early-stage (stages 0 and I) esophageal cancer samples ex vivo. Six marker bands were observed in the Raman spectra of cancerous tissues assessed by t-test, assignable to glycogen, collagen, and tryptophan, which presented decreased band intensity in the cancerous tissues. There were only slight spectral differences between cancerous and normal tissues, and the single chemometric technique was not sufficient to classify the tissue types accurately. Therefore, we performed partial least squares regression (PLSR) analysis and linear discriminant analysis (LDA) on only the significant wavenumbers assessed by the t-test to have statistically different Raman signal intensities. LDA based on Raman bands found in the t-test was able to predict the tissue types with 81.0% sensitivity and 94.0% specificity. Moreover, we found self-organization maps (SOMs) were suitable for the analysis of the present data, since they are among the artificial algorithms for neural networks whose learning process for pattern recognition resembles the one that is found in the brain. SOMs are a powerful tool to visualize data and the neural network statistical model may reveal the underlying patterns hidden in the obtained datasets. In practice, an initial map was built and was represented by a two-dimensional hexagon. Each map unit contains weight vectors randomly generated from a uniform distribution between the maximum and minimum values of variables in the data. Particularly, SNOMs show their power in pattern recognition involving noisy signals. AI-assisted diagnosis with Raman spectroscopy leads to early detection of the pathological changes that take place before the morphological changes and can result in a better outcome of the cancer therapy.