Multimodal Transformer-Based Preoperative Risk Assessment of Lymph Node Metastasis in Early Gastric Cancer Using Whole Slide Images of Endoscopic Biopsies.

backgroundAccurate preoperative assessment of lymph node metastasis (LNM) in early gastric cancer (EGC) is critical for choosing endoscopic or surgical treatment. Most computational pathology studies rely on resection specimens obtained after curative gastrectomy, which inherently limits their preoperative relevance. Our study developed a multimodal artificial intelligence (AI) framework to estimate LNM risk using preoperative whole slide images (WSIs) of endoscopic biopsies, aiming to assist clinicians in optimizing treatment strategies before the surgery and reducing unnecessary surgical risks for patients.

methodsWe developed a multimodal transformer-based model with a clinical-guided co-attention mechanism to fuse pathological features and clinical variables. To overcome the information scarcity inherent in fragmented biopsy specimens, we extracted mucosa and submucosa regions from WSIs of surgical tissues to serve as pseudo-biopsies. A curriculum learning (CL) paradigm was further employed to transition the model from idealized pseudo-biopsies to real-world biopsy samples.

resultsThrough the analysis of a dataset comprising 1,099 WSIs of surgical tissues and 189 WSIs of endoscopic biopsies, our framework achieved an area under the curve (AUC) of 0.759 on WSIs of surgical tissues and an AUC of 0.727 on preoperative WSIs of endoscopic biopsies.

conclusionOur findings demonstrate the feasibility of a multimodal AI framework tailored for endoscopic biopsy WSIs in preoperatively assessing LNM risk in T1 EGC. As an investigational clinical decision-support tool, this framework holds promising potential to assist clinicians in preoperative risk stratification, serving as a preliminary step toward optimizing prospective surgical management.