Researchers in Shanghai have built a multimodal artificial intelligence framework that estimates the risk of lymph node metastasis in early gastric cancer using only whole slide images of endoscopic biopsies — the tiny tissue samples obtained before any surgery takes place. The study, published in Annals of Biomedical Engineering, reports an area under the receiver operating characteristic curve (AUC) of 0.759 on whole slide images of surgical tissues and 0.727 on preoperative endoscopic biopsy slides, offering preliminary evidence that a meaningful predictive signal can be extracted from specimens that are barely two millimeters across.
The clinical stakes are considerable. Gastric cancer is the fourth most prevalent malignancy worldwide and the third leading cause of cancer death, and early gastric cancer — defined pathologically as tumor confined to the mucosa (T1a) or submucosa (T1b) — carries dramatically different metastatic risks depending on invasion depth: roughly 2% for mucosal cases and about 18% for submucosal cases. That gradient drives treatment decisions. When lymph node risk is negligible, patients can undergo organ-preserving endoscopic resection procedures such as endoscopic submucosal dissection or endoscopic mucosal resection; when risk features emerge, radical gastrectomy with lymph node dissection follows.
The problem, the authors argue, is that current risk tools overtrigger surgery. Widely adopted systems such as the eCura score, which stratifies risk from pathological features of resection specimens, lead many patients to additional radical gastrectomy — yet the actual rate of lymph node metastasis found in those follow-up surgeries is only 5 to 10%. In other words, up to 95% of these invasive operations may be unnecessary, exposing patients to surgical morbidity without demonstrable benefit. A more precise preoperative estimate, available before any resection, could spare many of those patients.
Predicting metastasis from endoscopic biopsies, however, is fundamentally harder than from resected stomachs. Biopsy samples measure approximately 0.2 centimeters in diameter and offer only fragmentary glimpses of the tumor microenvironment, while surgical specimens encompass the entire primary lesion plus peritumoral stroma, lymphovascular structures and adjacent normal tissue. Most prior computational pathology studies have sidestepped this difficulty by training on post-gastrectomy slides — a setting that inherently limits preoperative relevance. Previous models illustrate the drop-off: one multimodal multiple-instance-learning approach reported an AUC of 0.877 on all-stage surgical specimens but fell to 0.725 on biopsy slides.
To overcome the scarcity of real biopsy material, the team devised an unsupervised segmentation strategy that extracts mucosa and submucosa regions from whole slide images of surgical tissues, creating what they call pseudo-biopsies. The pipeline projects patch-level embeddings, generated by the Prov-GigaPath pathology foundation model — a Vision Transformer Giant encoder with 1.3 billion parameters trained on 1.3 billion image tiles — into a two-dimensional manifold using t-SNE, then applies K-means clustering to separate target mucosal-submucosal patches from everything else. The correct cluster is identified automatically by maximizing the probability density of blue-channel stain intensity within a characteristic range, followed by Laplacian edge enhancement and connected-component analysis to remove fragmented artifacts and redundant muscularis tissue.
The quality control was rigorous. An expert pathologist reviewed the segmentation outputs across all 1,099 surgical whole slide images against predefined criteria, rating 985 slides (89.6%, 95% CI 87.7–91.3%) as excellent and 54 (4.9%) as acceptable, for an overall clinical usability rate of 94.5%. On a subset of 20 fully annotated slides, the automated method achieved a mean Dice similarity coefficient of 0.846, a mean intersection-over-union of 0.727 and a mean precision of 0.926 against the expert ground truth.
On top of these features sits a dual-stream Transformer architecture fused by a clinical-guided co-attention mechanism. Instead of simply concatenating clinical variables with image features, the co-attention layer uses patient-specific data — sex, age, tumor location and tumor size — as queries that actively guide the aggregation of patch-level pathological features, assigning higher importance weights to instances with high diagnostic relevance given the clinical profile. Training followed a two-stage curriculum learning paradigm: the model first learned from the idealized, spatially contiguous pseudo-biopsies, then was fine-tuned on real, fragmented endoscopic biopsy slides to adapt to the target domain.
The cohort comprised 637 patients with pathologically confirmed T1 gastric cancer who underwent subtotal or total gastrectomy at Zhongshan Hospital of Fudan University between January 2008 and December 2022, yielding 1,099 surgical whole slide images and 189 paired endoscopic biopsy slides. Data were split at the patient level, with 20% held out as an untouched independent test set and five-fold stratified cross-validation used for tuning. In head-to-head comparisons under identical data splits and feature pipelines, the framework outperformed established multiple-instance-learning baselines including Deep Sets, Attention MIL, DSMIL, CLAM, TransMIL and DTFD-MIL, beating the strongest baseline, DTFD-MIL, by 3.1 percentage points (AUC 0.759 versus 0.728) on the surgical cohort and surpassing all baselines on the biopsy cohort. Ablation experiments underscored each component’s contribution: removing the MIL Transformer dropped biopsy-cohort AUC to 0.612, dropping curriculum learning reduced it to 0.635, and omitting co-attention lowered it to 0.686. On an external TCGA-STAD validation set of just 50 patients, the model reached an AUC of 0.751, though the wide 95% confidence interval (0.613–0.892) reflects the small, imbalanced sample and warrants caution.
The model’s attention maps also proved biologically informative. Heatmaps consistently concentrated high attention on tumor tissue and tumor-infiltrating lymphocytes while ignoring normal glandular epithelium, and the model selectively prioritized tertiary lymphoid structures — immune cell aggregates within the tumor microenvironment that harbor effector populations such as CD8-positive T cells. In a sub-cohort of 290 histologically verified cases, regions of higher model attention showed significantly greater tertiary lymphoid structure density than low-attention regions (Wilcoxon signed-rank test, p = 0.0017, r = 0.839), and high-attention-region density was significantly higher in metastasis-positive than metastasis-negative cases across the full cohort (Mann–Whitney U test, p < 0.0001, r = 0.322), the latter modest effect size attributed to the low prevalence of these structures in immunologically quiescent T1 tumors.
The authors are explicit that this remains an investigational decision-support tool, not a clinically validated test. At the observed lymph node metastasis prevalence of 13.2%, the biopsy operating point — sensitivity 0.733 and specificity 0.682 — yields a projected positive predictive value of 26.0% and negative predictive value of 94.4%, meaning roughly 32 of every 100 metastasis-negative patients would be incorrectly flagged as high risk. Meaningful translation, they note, would require decision-specific thresholds targeting sensitivity of at least 90–95% and a negative predictive value of at least 95%, followed by assessments of calibration and net benefit in larger, multi-institutional, prospectively collected cohorts. Still, the study demonstrates the feasibility of extracting preoperative metastasis risk from the least invasive specimen available — a first step toward personalized treatment planning at the moment of initial diagnosis, rather than after the stomach has already been removed.
Subject of Research: Multimodal AI prediction of lymph node metastasis in early gastric cancer from preoperative endoscopic biopsy whole slide images
Article Title: Multimodal Transformer-Based Preoperative Risk Assessment of Lymph Node Metastasis in Early Gastric Cancer Using Whole Slide Images of Endoscopic Biopsies
Article References: Niu, X., Wang, S., Yu, J., Tian, M., & Tang, Z. (2026). Multimodal Transformer-Based Preoperative Risk Assessment of Lymph Node Metastasis in Early Gastric Cancer Using Whole Slide Images of Endoscopic Biopsies. Annals of Biomedical Engineering. https://doi.org/10.1007/s10439-026-04380-7
Image Credits: AI Generated
DOI: 10.1007/s10439-026-04380-7
Keywords: early gastric cancer, lymph node metastasis, whole slide imaging, artificial intelligence, transformer, multiple instance learning, endoscopic biopsy, curriculum learning, computational pathology, Prov-GigaPath, clinical decision support, tertiary lymphoid structures
News Source: Nathaniel Bowman. (October 23, 2026). AI Reads Preoperative Endoscopic Biopsies to Gauge Lymph Node Spread in Early Gastric Cancer. Scienmag.
