Combining Raman spectroscopy and machine learning to assist early diagnosis of gastric cancer

Abstract

Gastric cancers, with gastric adenocarcinoma (GAC) as the most common histological type, cause quite a few of deaths. In order to improve the survival rate after GAC treatment, it is important to develop a method for early detection and therapy support of GAC. Raman spectroscopy is a potential tool for probing cancer cell due to its real-time and non-destructive measurements without any additional reagents. In this study, we use Raman spectroscopy to examine GAC samples, and distinguish cancerous gastric mucosa from normal gastric mucosa. Average Raman spectra of two groups show differences at 750 cm−1, 1004 cm−1, 1449 cm−1, 1089–1128 cm−1, 1311–1367 cm−1 and 1585–1665 cm−1, These peaks were assigned to cytochrome c, phenylalanine, phospholipid, collagen, lipid, and unsaturated fatty acid respectively. Furthermore, we build a SENet-LSTM model to realize the automatic classification of cancerous gastric mucosa and normal gastric mucosa, with all preprocessed Raman spectra in the range of 400–1800 cm−1 as input. An accuracy 96.20% was achieved. Besides, by using masking method, we found the Raman spectral features which determine the classification and explore the explainability of the classification model. The results are consistent with the conclusions obtained from the average spectrum. All results indicate it is potential for pre-cancerous screening to combine Raman spectroscopy and machine learning.