Northeastern finds vision transformer detects more precancerous lesions in single-dataset test

A vision transformer identified precancerous colorectal lesions with 70% sensitivity, compared with 40% for the strongest competing convolutional network, Northeastern University reported October 19. The finding points to a possible improvement in recognizing tissue that can resemble harmless polyps. But the comparison covered only one public dataset, and researchers say broader validation is needed.

The medical challenge is a subtle distinction on biopsy slides. Sessile serrated adenomas are precancerous lesions in the colon’s inner lining. They can look much like hyperplastic polyps, which are benign, according to pathologist Christopher Owens, an associate medical director at Quest Diagnostics. Distinguishing the two requires close attention to small differences in the structure of the tissue’s cells.

Owens described earlier detection as a way to prevent more cases from progressing to colorectal cancer. Many patients in the disease’s early stages have no symptoms, he said. That gives the distinction between a harmless polyp and a precancerous lesion practical importance: the task is not simply finding an abnormality, but recognizing which kind of tissue it is.

Researchers at Northeastern’s Amal Lab for Precision Medicine compared three vision transformers with three convolutional neural networks, or CNNs. They used images of more than 3,000 colorectal polyps from a public Dartmouth-Hitchcock Medical Center dataset. The collection included 1,545 benign hyperplastic polyps and 630 sessile serrated adenomas; those were not the only polyps in the dataset.

A wider view of the tissue

Sensitivity measures how well a tool identifies positive cases—in this comparison, the precancerous lesions. The stronger transformer result therefore concerns the cases researchers wanted the systems to catch, rather than overall accuracy alone. The best transformer also led on that broader measure, although its sensitivity still fell short of identifying every precancerous case.

Saeed Amal, a Northeastern bioengineering professor, offered a possible explanation for the gap. Vision transformers can examine relationships across an image, while CNNs tend to focus more locally, analyzing it segment by segment. He compared the difference to viewing the sweep of the night sky rather than studying an individual star through a telescope. That broader view may help explain the transformer’s performance, he said.

From a model comparison to patient care

Amal acknowledged that testing on a single dataset limits the finding and said additional datasets are needed to validate it. The result compares AI systems with one another; it does not demonstrate that the transformer finds lesions pathologists cannot see. Nor does that comparison establish that using it will prevent more cancers.

A related clinical concern comes from AI-assisted colonoscopy, a different setting from examining biopsy tissue. In its March 2025 guideline, the American Gastroenterological Association said the technology’s long-term effect on reducing cancer cases remained unknown and did not endorse widespread use. The association warned that detecting more low-risk polyps could mean more frequent, costly colonoscopies with little benefit.

Amal said his team has begun working with clinicians to test the systems. He described prospective studies—evaluations as clinical work happens—with two complementary roles:

• Pathologists would verify the AI’s findings by examining tissue, checking whether its classifications hold up in the diagnostic work the system is intended to support.
• Gastroenterologists would assess how those findings could guide patient care. Together, the clinicians would help evaluate accuracy, usefulness and how the algorithms fit into real clinical workflows.