Gastric cancer (GC) is a leading cause of cancer death globally. Early diagnosis is critical, as the five-year survival rate drops from up to 99% for early-stage GC (EGC) to under 30% for advanced stages. While standard endoscopy screening has lowered mortality in some countries, it is generally too expensive and invasive for large-scale public screening programs.
To address this, researchers developed GRAPE (GC risk assessment procedure with AI), a deep-learning model that analyzes routine noncontrast computed tomography (CT) scans to identify high-risk patients. Noncontrast CT is a low-cost and widely available imaging protocol, making it ideal for opportunistic screening.
The GRAPE AI model showed high performance, achieving an AUC of 0.927 in external validation. Its detection rate exceeded 90% for advanced GC and was approximately 50% for EGC. In a reader study, GRAPE alone (AUC 0.92) outperformed all radiologists (AUC 0.76–0.85), though AI assistance significantly improved radiologist accuracy. In a massive real-world opportunistic screening study of over 78,000 participants, GRAPE demonstrated high sensitivity (88.7%−91.9%) and successfully predicted GC in over 63% of patients using CT scans up to 6 months before diagnosis.
While GRAPE is not intended to replace endoscopic evaluation, it offers a highly accurate, non-invasive, and cost-effective method for mass GC screening that can prioritize high-risk populations for follow-up endoscopy, thereby enhancing screening compliance and ultimately reducing GC mortality.