Alibaba’s Damo Academy, in partnership with Zhejiang Cancer Hospital, has launched Grape AI, a model built to detect stomach cancer from non-contrast CT scans.
As a first-of-its-kind system for gastric cancer screening, Grape analyzed data from nearly 100,000 participants across a multicenter study. Moreover, it achieved a whopping 85.1 percent sensitivity and 96.8 percent specificity, surpassing human radiologists.
The model successfully identified tumors that clinicians had missed, sometimes months before a traditional diagnosis.
Despite the low cost and widespread availability of CT scans, early gastric cancer could not be detected with traditional 3D imaging owing to small mucosal lesions and moving anatomy.
Grape overcomes these challenges by using deep learning, which was trained on the most comprehensive annotated dataset of stomach cancer CT scans ever gathered. The model not only recognizes common tumors, but it also marks early-stage cancers hidden within the gastric lining, which previously required contrast-enhanced imaging.
During regional deployments in Anhui and Zhejiang provinces, Grape was able to detect 24.5% of high-risk patients in one hospital and 17.7% in another, even though majority of these patients displayed no symptoms.
In one illustrative instance, the AI identified a patient’s malignancy six months before the clinical diagnosis came in. These findings indicate that Grape has the potential to substantially expedite the diagnosis process and mitigate care delays.
Gastric cancer ranks among the top four causes of cancer-related deaths globally. In China, mortality hovers near 260,000 annually, with five-year survival rates for early detection less than 30 percent rising to over 95 percent if caught early.
However, fewer than 30 percent of high-risk patients undergo invasive endoscopy screens. Grape’s use of standard CT scans could bypass these barriers and broaden screening accessibility.
Grape follows Alibaba’s Damo Panda, an AI tool for pancreatic cancer that received FDA Breakthrough Device status in April. Damo Panda demonstrated 92.9 percent sensitivity and 99.9 percent specificity in detecting early pancreatic cancer on non-contrast CT scans, surpassing radiologist performance by 34 percent.
Together, these models highlight Alibaba’s push to embed AI in routine medical imaging for early disease detection.