Alibaba released its fourth cancer screening AI model. The DAMO EAGLE, an esophageal cancer screening AI model developed by Alibaba DAMO Academy in collaboration with Sichuan Cancer Hospital and the Sun Yat-sen University Cancer Center, can identify esophageal cancer (including early-stage and precancerous malignant lesions) from plain CT scans without the need for endoscopy or contrast agents. It has been validated on 80,000 cases across three countries, and the related paper was published on September 22 in the international top-tier journal Nature Medicine.

image.png

Cracking the 'Impossible' Task, Sensitivity 90%, Specificity 99.2%

The incidence and mortality rate of esophageal cancer in China account for nearly half of the global total, with most patients being diagnosed at a late stage and missing the chance for curative treatment. If detected early, most patients can be cured through minimally invasive endoscopic surgery, with a five-year survival rate exceeding 95%. The current standard screening method is upper gastrointestinal endoscopy, which is somewhat invasive and complicated, making it difficult to promote widely. However, plain chest CT is low-cost and well-accepted, and its scanning range naturally covers the esophagus—if it can identify esophageal cancer risk while screening for lung cancer, it would be highly efficient. The challenge lies in the fact that the esophagus is a long, easily collapsed hollow organ, and early lesions are hidden in the epithelial and mucosal layers, compounded by the movement of the heart and major blood vessels, making it difficult for radiologists to identify from plain CT scans.

Yao Jiawen, an algorithm expert from DAMO Academy, explained that the team innovatively matched endoscopic reports with accurate lesion locations on enhanced CT images to plain CT images, successfully training the AI to identify early lesions that are hard for the human eye to detect. According to the paper, under opportunistic screening scenarios, DAMO EAGLE achieves a sensitivity (no missed diagnoses) of 90% for esophageal cancer, and a sensitivity of 52.5% for signals that are extremely weak in precancerous lesions. In real-world scenarios, its specificity (no false positives) reaches as high as 99.2%, and its performance remains unchanged on low-dose CT, offering the potential to seamlessly integrate into existing lung cancer screening processes without adding patient burden.

Completing the 'Plain CT + AI' Multi-Check Route

Wang Qifeng, a chief physician from the Radiotherapy Department of Sichuan Cancer Hospital, stated that due to habits such as consuming hot, salty foods, esophageal cancer is prevalent in western China, especially in Sichuan. Due to limited medical resources and patient compliance, there is a significant gap in endoscopic screening. "AI is not replacing endoscopy, but rather identifying high-risk individuals who then undergo endoscopy, making the examination more accurate and significantly improving detection rates," he said. More importantly, it moves the prevention and control关口 forward, providing more opportunities for early detection.

With the continuous release of AI models for pancreatic cancer, gastric cancer, colorectal cancer, esophageal cancer, and aortic dissection, DAMO Academy has successfully implemented the original technical route of "plain CT + AI" for multi-checking. It has published a total of five papers in Nature Medicine, with the potential to detect the top seven cancers causing the most deaths in China through a single plain CT scan.