Lab tests: AI transformer flags 70% of sessile serrated colon lesions

Dr. Christopher Owens, an associate medical director at Quest Diagnostics in Marlborough, Massachusetts, has spent his career trying to catch precancerous colorectal polyps earlier. The hard cases are the sessile serrated adenomas, or SSAs. Under the microscope they look almost identical to a hyperplastic polyp, which is completely benign. A new study from Northeastern University suggests an AI model could help tell them apart: in lab tests, the best vision transformer detected 70% of precancerous lesions.

Vision transformers are the architecture behind AI chatbots. They analyze how different parts of an image relate to one another, whereas the older convolutional neural networks, or CNNs, examine an image segment by segment. Saeed Amal, a Northeastern professor of bioengineering, compares the CNN to an astronomer studying one star through a telescope, and the transformer to an eye that takes in the whole night sky. The team compared three models of each type on more than 3,000 polyp images from a public data set provided by Dartmouth-Hitchcock Medical Center, which included 1,545 hyperplastic polyps and 630 SSAs. The best transformer's sensitivity was 70%, against 40% for the CNN, and its overall accuracy was 86%, against 80%.

The limits are clear, and the authors name them. The models were analyzed on one data set only, and further testing on multiple data sets is required. The American Gastroenterological Association, in a March 2025 clinical guideline, said the long-term effect of AI-assisted colonoscopy on cancer cases is "as of yet unknown", and does not endorse its widespread use. The association also warned that such systems tend to flag low-risk polyps, which could mean more frequent and costly colonoscopies with little benefit. Its expert Shahnaz Sultan put the bar plainly: "Right now, AI is detecting easy-to-detect lesions. This is version 1.0."

For patients and pathologists, the gain would be concrete if the results hold. A pathologist could get a second look at a lesion the eye alone might miss, and an SSA caught earlier means a precancerous lesion removed before it can become cancer. The team has begun working directly with clinicians: pathologists to verify findings in tissue and gastroenterologists to judge how they guide care, and plans prospective studies that would test the algorithms in real time, within clinical workflows. No date for those trials, nor for any clinical use, appears in the sources consulted.

Leon Toussaint, scientific manager at the National Foundation for Cancer Research, praised the work for showing "where AI succeeds and where it fails in imbalanced real-world data". That is the test the next studies will have to pass before the gap seen in the lab can reach a patient's colon.