A collaboration of leading Chinese research institutions has developed an artificial intelligence-based method called GRAPE, demonstrating high accuracy in detecting gastric cancer from routine noncontrast CT scans.

Gastric cancer ranks among the most lethal malignancies worldwide, particularly in Asian populations. In China, Japan, and Korea, nearly three-quarters of new diagnoses and deaths cluster each year, reflecting limited early detection and treatment barriers. Endoscopy remains the benchmark for diagnosis, allowing clinicians to visualize the gastric mucosa and collect biopsies for confirmation.

National screening programs in Japan and Korea have raised survival rates through widespread endoscopic exams. Many countries lack the resources to deploy such strategies, and the invasiveness of the procedure, along with social perception, further reduce compliance rates. Serological screening offers a less intrusive alternative but achieves only marginal gains over population-wide gastroscopy.

Low compliance, limited detection rates, and prohibitive costs have left an urgent demand for affordable, noninvasive methods to pinpoint high-risk individuals before cancer advances beyond curative stages.

In the study, “AI-based large-scale screening of gastric cancer from noncontrast CT imaging,” published in Nature Medicine, researchers developed GRAPE (Gastric Cancer Risk Assessment Procedure with Artificial Intelligence) to identify gastric cancer (GC) patients through deep learning analysis of noncontrast computed tomography (CT) scans.

Researchers trained GRAPE using data from two centers in China, encompassing 3,470 GC cases and 3,250 non-cancer cases. Following its development, GRAPE underwent extensive validation with an internal cohort of 1,298 cases, achieving a sensitivity of 85.1%, specificity of 96.8%, and an area under the curve (AUC) of 0.970. Validation on an external cohort from 16 centers with 18,160 cases confirmed stable performance, yielding an AUC of 0.927 and sensitivity and specificity scores of 81.7% and 90.5%, respectively.

Further evaluation came through reader studies involving 13 radiologists interpreting 297 scans. GRAPE consistently outperformed human readers, significantly improving sensitivity by 21.8% and specificity by 14.0%, particularly for early-stage GC cases.

Even after radiologists re-evaluated scans with GRAPE assistance following a washout period, the AI maintained superior accuracy, underscoring its potential role as a robust diagnostic support tool.

Validation in real-world settings involved analysis of 78,593 consecutive noncontrast CT scans collected between 2018 and 2024 from one comprehensive cancer center (Zhejiang Cancer Hospital) and two regional hospitals (Fenghua People’s Hospital and Pingyang People’s Hospital).

GRAPE identified high-risk individuals effectively, showing GC detection rates of 24.5% in Fenghua and 17.7% in Pingyang. Approximately 40% of these detected cases lacked previous abdominal symptoms. In the Zhejiang Cancer Center cohort, GRAPE detected cancer at a rate of 12.1%, even identifying tumors months ahead of clinical diagnosis in patients followed for other cancers.

GRAPE integrates both tumor segmentation and patient-level classification within a single deep-learning framework. Initially, GRAPE identifies the stomach area on the full CT image, then crops this region to detect tumors and simultaneously classify the patient as having or not having GC.

Visual analysis demonstrated clear delineation of tumor areas, aligning well with GRAPE’s predictions and enabling radiologists to quickly interpret results.

While GRAPE achieved strong overall detection rates, its sensitivity for the earliest-stage cancers remained limited. The system identified approximately 50% of early stage (T1) GCs in validation cohorts, reflecting the challenges of detecting small or subtle lesions on noncontrast CT scans.

By comparison, early stage GCs are precisely what endoscopic examination excels at detecting, as it enables direct visualization of minor mucosal changes and allows tissue sampling for confirmation.

Researchers acknowledge further refinement and testing are necessary, specifically a large prospective screening trial to assess GRAPE’s real-world efficacy and optimize its sensitivity for early-stage cancers.

Plans include expanding training data to encompass earlier-stage cancers and incorporating detailed pathological insights. Researchers also suggest procedural improvements, such as stomach distension before imaging, to enhance early-stage detection.

GRAPE presents a substantial advancement in large-scale GC screening, offering significant potential for improving early diagnosis rates through a more accessible, cost-effective, and noninvasive method.

© 2025 Science X Network