Intelligent Diagnosis of Pancreatic Biopsy From Endoscopic Ultrasound-Guided Fine-Needle Aspiration Via Stimulated Raman Histopathology

Endoscopic ultrasound–guided fine-needle aspiration (EUS-FNA) has become one of the most important preoperative diagnostic methods for pancreatic tumors, but it often faces challenges of redundant sampling from patients and complex tissue processing that hinders timely diagnosis. Intraoperative rapid on-site evaluation is an auxiliary diagnostic technique that helps assess sample quality in real time, but it heavily depends on pathologists and involves subjectivity and complex procedures. Here, we developed a rapid and label-free approach for intraoperative histology on EUS-FNA specimen via deep learning–based stimulated Raman scattering microscopy, aimed at replacing rapid on-site evaluation and providing a more efficient and objective diagnostic approach. Fresh pancreatic EUS-FNA tissues were imaged with stimulated Raman scattering and compared with hematoxylin and eosin staining to identify key histologic features. Using images from 76 patients, a convolutional neural network model was established to identify benign, malignant, and nondiagnostic areas, achieving a validation accuracy >96% on an external test set of 33 cases. Furthermore, gradient-weighted class activation mapping was able to highlight histologic profiles within individual biopsy. Our approach has potential application in efficient intraoperative assessment of pancreatic biopsy through EUS-FNA.