Accurate and Rapid Detection of Peritoneal Metastasis from Gastric Cancer by AI-assisted Stimulated Raman Cytology

Abstract

AbstractPeritoneal metastasis (PM) is the most common form of distant metastasis and one of the leading causes of death in gastric cancer (GC). For locally advanced GC, clinical guidelines recommend peritoneal lavage cytology for intraoperative PM detection. Unfortunately, current peritoneal lavage cytology is limited by low sensitivity (<60%). Here we established the stimulated Raman cytology (SRC), a chemical microscopy-based intelligent cytology. By taking advantages of stimulated Raman scattering in label-free, high-speed, and high-resolution chemical imaging, we firstly imaged 53951 exfoliated cells in ascites obtained from 80 GC patients (27 PM positive, 53 PM negative), at the Raman bands corresponding to DNA, protein, and lipid, respectively. Then, we revealed 12 single cell features of morphology and composition that were significantly different between PM positive and negative specimens, including cellular area, lipid protein ratio, etc. Importantly, we developed a single cell phenotyping algorithm to further transform the above raw features to feature matrix. Such matrix was crucial to identify the significant marker cell cluster, the divergence of which was finally used to differentiate the PM positive and negative. Compared with histopathology, the gold standard of PM detection, our SRC method assisted by machine learning classifiers could reach 81.5% sensitivity, 84.9% specificity, and the area under receiver operating characteristic curve of 0.85, within 20 minutes for each patient. Such remarkable improvement in detection accuracy is largely owing to incorporation of the single-cell composition features in SRC. Together, our SRC method shows great potential for accurate and rapid detection of PM from GC.