Accurate and rapid molecular cytology with transformer-based stimulated Raman microscopy

Owing to the minimal invasiveness, cytology is an indispensable technique in the routine pathology. However, traditional cytology only enables the low sensitivity (50%-60%) and high time-consuming for the diagnosis. Our previous study demonstrated that stimulated Raman molecular cytology (SRMC), which is label-free, faster, and noninvasive, provides additional composition information leading to higher diagnostic accuracy around 85%. However, current AI-assisted SRMC generally involves cell segmentation and feature extraction steps, which may involve issues of the artifacts. Recently, various methods for global feature analysis, such as Transformer and CNN, are capable of preserving both global and local information. Therefore, we propose an end-to-end Transformer hybrid model combining the advantages of both Transformer and CNN to analyze stimulated Raman cytology images for accurate and rapid peritoneal metastasis (PM) diagnosis of gastric cancer (GC). The Transformer hybrid model can enhance the Transformer’s global modeling ability simultaneously with the local guidance from CNN features. To evaluate the performance of this Transformer method, we collected 816 stimulated Raman cytology images from 80 locally advanced gastric cancer patients, with 36 PM positive and 44 PM negative. The Transformer method could reach 88.89% sensitivity, 86.36% specificity, and an AUC of 0.903 with leave-one-out cross-validation for 80 patients. Compared with traditional cytology, the false negative rate of our label-free stimulated Raman cytology reduces by about 30-50%. Together, our Transformer approach demonstrates the potential for accurate and rapid PM diagnosis based on exfoliated cytology.