Abstract

Introduction: Gastric cancer (GC) is a heterogeneous disease, with Lauren classification providing a framework to assign diffuse and intestinal subtypes, informing prognosis and therapy. Traditional subtype assignment relies on pathologist review of hematoxylin and eosin (H&E)-stained slides, leading to inter-observer variability and scalability challenges. We developed a deep learning classifier to automate Lauren subtype assignment on real-world H&E images.

Methods: We analyzed de-identified H&E-stained whole slide images (WSI) from biopsies and resections of 2974 GC patients (3160 samples) from the Tempus real-world database. Samples with pathologist-assigned labels (n=399 diffuse; n=238 intestinal) were used for classifier training. WSI were preprocessed into tissue tiles and tile embeddings were extracted using the H-optimus-0 pathology foundation model. An additive attention-based multiple instance learning model was trained with cross-entropy loss weighted by class prevalence. Data were split 80/20 for development/holdout, with 5-fold cross-validation for model tuning and selection, and ensembled predictions from 5 cross-validation models were used to assign subtypes. An operating point was selected for ~90% positive predictive value (PPV) on the holdout set for each class. Real-world overall survival (rwOS; time from first-line therapy to death) was assessed in patients with available data (31%).

Results: The model achieved a robust performance (AUC 0.93, 95% CI: 0.88-0.98) on the holdout set. With PPV-optimized thresholds, previously unlabeled samples (n=2523) were assigned by the model as diffuse (n=1321, 52.4%), intestinal (n=749, 29.7%), or indeterminate (n=453,17.95%). For pathologist-assigned samples, diffuse cases had worse median OS (13.3 months, 95% CI: 11.5-15.8) than intestinal (22 months, 95% CI: 15.1-29.8; p=6.2e-4). For classifier-assigned samples, diffuse cases had a shorter median OS (12.6 months, 95% CI: 10.8-15.7) than intestinal (15.3 months, 95% CI: 12.2-17.3; p=0.95). CDH1 mutations were found in 30.3% of pathologist-labeled and 23.7% of classifier-assigned diffuse tumors, but were rare in intestinal tumors (1.3%, 1.1%). RHOA mutations were present in 8.5% of pathologist-labeled and 8.3% of classifier-assigned diffuse tumors, versus 2.5% in intestinal tumors for both groups. Other histologies predominantly aligned with model predictions: signet ring cell carcinoma was predicted diffuse, while tubular, papillary, and mucinous adenocarcinomas were predicted intestinal.

Conclusions: This deep learning classifier can accurately assign Lauren subtypes in GC from real-world H&E-stained WSI, reducing manual review and variability. Model predictions align with known clinical and molecular differences between subtypes, supporting standardization of Lauren classification and enabling large-scale studies of GC.

VIEW THE PUBLICATION