Large-Scale Gastric Cancer Screening Achieved with Plain CT + AI: Zhejiang Tumor Hospital and DAMO Academy Develop GRAPE, Published in Nature Medicine}
Researchers from Zhejiang Tumor Hospital and DAMO Academy developed GRAPE, an AI-powered system using plain CT scans for early gastric cancer detection, published in Nature Medicine, demonstrating high accuracy.
Author: Sun Qiushi, a PhD student at the School of Data Science, University of Hong Kong, with a master's from NUS Data Science. Focuses on Computer-Using Agents and Code Intelligence, with multiple publications at ACL, EMNLP, ICLR, COLING, and more. The OS-Copilot team previously released OS-Atlas, OS-Genesis, and SeeClick, widely used in academia and industry.
In China, over 30% of gastric cancer (GC) patients are diagnosed at an advanced stage, missing the opportunity for surgery. Early detection is crucial to reduce mortality.
However, resource limitations, low compliance, and suboptimal endoscopic screening rates in high-incidence areas make large-scale screening challenging.
Scientists have turned to AI to analyze routine clinical plain CT scans for cancer detection. A joint team from Zhejiang Tumor Hospital, Hangzhou Medical College, and DAMO Academy developed GRAPE (Gastric Cancer Risk Assessment Procedure with Artificial Intelligence), which uses deep learning on non-contrast CT images to identify early gastric cancer.
Remarkably, GRAPE can detect early gastric cancer solely by analyzing standard hospital plain CT images.
The study titled “AI-based large-scale screening of gastric cancer from noncontrast CT imaging” was published on June 24, 2025, in the top medical journal Nature Medicine.
To clarify, prognosis of gastric cancer is directly related to staging. As stage increases from T1 to T4, tumors invade deeper layers, and once reaching late stages or invading the serosa, surgery is no longer curative.
Statistics show that T1 stage patients have a 97-99% cure rate, while T4 stage patients have an average survival of only 10-12 months, regardless of treatment. Early detection and treatment significantly improve survival, which underscores the value of GRAPE.
About GRAPE
GRAPE is a deep learning framework designed to analyze 3D non-contrast CT scans for detecting and segmenting gastric tumors. It was trained on a dataset of 3,470 gastric cancer cases and 3,250 non-cancer cases, producing pixel-level tumor masks and classification scores distinguishing GC from NGC.
The model employs a two-stage approach: first, a segmentation network locates the stomach in the full CT scan, generating a mask; then, the masked region is processed by a dual-branch network for tumor detection and classification, identifying GC-positive or negative cases.
Performance Validation
Validation on internal datasets (1,298 cases, AUC=0.970) and external multi-center datasets (18,160 cases, AUC=0.927) demonstrated GRAPE’s high accuracy. Its AUC (0.757-0.79) surpasses traditional models based on clinical and serological data, comparable to liquid biopsy methods.
Subgroup analysis shows GRAPE’s sensitivity for early gastric cancer (~50%) and over 90% for T3/T4 stages.
Reader studies comparing GRAPE with radiologists reveal that GRAPE significantly outperforms radiologists, with a 21.8% increase in sensitivity and 14.0% in specificity, especially for early-stage GC.
Real-world case studies in a comprehensive cancer center and two regional hospitals involving 78,593 non-contrast CT scans showed GRAPE’s ability to identify high-risk groups, with detection rates of 24.5% and 17.7%, including early-stage cases.
GRAPE also detects cases missed by radiologists, enabling earlier diagnosis during follow-ups for other diseases. It performs well on pre-diagnosis scans up to six months before confirmed diagnosis.
Significance
While not intended to replace endoscopy, GRAPE offers a valuable alternative for symptomatic patients unwilling to undergo initial endoscopy. Its low cost (~200 RMB) compared to traditional screening (~3000 RMB) makes it suitable for large-scale screening.
The DAMO team emphasizes that this “one-scan multi-screen” approach can effectively address diagnostic blind spots caused by specialist fatigue and improve detection accuracy across the entire CT imaging process. AI’s global screening capability can significantly enhance diagnostic sensitivity and reduce workload.
In summary, GRAPE demonstrates strong potential for large-scale gastric cancer screening, providing an effective solution for early cancer detection and safeguarding human health.
Paper link: https://www.nature.com/articles/s41591-025-03785-6