视觉 Transformer 模型检出结直肠癌前病变敏感性 70%,高于 CNN 的 40%
AI is helping pathologists find harmful lesions the eye can't see
美国东北大学 Amal 实验室在《Lecture Notes in Computer Science》发表研究,视觉 Transformer 模型在结直肠息肉病理图像中检出癌前病变(无蒂锯齿状腺瘤)的敏感性为 70%,高于表现最好的 CNN 的 40%。
As a pathologist, Dr. Christopher Owens has spent his career trying to improve his methods of catching precancerous colorectal polyps in patients. If he could catch cases earlier, fewer people would develop colorectal cancer, he said.
An estimated 1.9 million people around the world are diagnosed with colorectal cancer every year, according to the World Health Organization.
Many in the early stages of the disease show no symptoms, noted Owens, an associate medical director at Quest Diagnostics, a clinical lab in Marlborough, Massachusetts, that provides cancer screening tests.
But one early warning sign that a patient may develop colon cancer is the presence of sessile serrated adenomas, or SSAs, in their tissue. SSAs are precancerous lesions found in the inner lining of the colon. However, these lesions can be hard to detect in certain cases because they are hard to spot on biopsy slides, Owens said.
"They look very much like a polyp called a hyperplastic polyp, which is completely benign," he said. "We're looking for subtle clues in the architecture of the cells themselves to determine which type of polyp it is."
Vision transformers offer a detection advantage
Northeastern University findings, published in Lecture Notes in Computer Science, show that a class of screening models equipped with the same foundational technology found in AI chatbots could be key in helping doctors like Owens spot those polyps more easily.
Specifically, researchers at Northeastern's Amal Lab for Precision Medicine have found that vision transformer models—a type of AI technology that can understand how different parts of an image relate—may be more equipped to identify precancerous lesions than an older class of AI models known as convolutional neural networks, or CNNs.
"This means we will be able to intervene earlier to stop precancerous lesions from transitioning into cancer," said Saeed Amal, Northeastern professor of bioengineering.
AI screening faces questions about clinical benefit
While AI has begun to be used in the medical space, it has received some pushback.
In March 2025, the American Gastroenterological Association issued a clinical guideline regarding the use of AI-assisted technologies in colonoscopies.
It noted that while AI is promising, its long-term impact in "reducing cancer cases is as of yet unknown." The organization does not endorse the widespread use of the technology.
One issue with AI-assisted detection systems is that they have a tendency to identify "low-risk polyps," according to the American Gastroenterological Association. That in turn could result in more frequent and costly colonoscopies with little benefit, the association said.
"If AI is going to be impactful, it needs to be better than the human eye," Dr. Shahnaz Sultan, one of the authors of the association's guidelines, said in an accompanying statement. "Right now, AI is detecting easy-to-detect lesions. This is version 1.0. Before we can recommend everyone use AI, we need version 4.0, where it helps detect polyps that are truly difficult to find."
A broader view improves lesion detection
For the Northeastern study, the researchers took three vision transformer models and three CNN models and had them analyze images of more than 3,000 colorectal polyps sourced from a public data set provided by the Department of Pathology and Laboratory Medicine at Dartmouth-Hitchcock Medical Center. The data set included 1,545 hyperplastic polyps, which are benign, and 630 sessile serrated adenomas, which are precancerous.
The researchers found that the best-performing vision transformer achieved a 70% sensitivity rate when it came to identifying cases of precancerous lesions, compared to the 40% sensitivity rate of the top-performing CNN. Sensitivity measures a tool's ability to identify positive cases.
The vision transformer model also achieved a greater overall accuracy at 86% compared to the CNN's 80%, according to the research.
One reason vision transformer models may perform better is that they can analyze an image holistically, while CNNs tend to analyze an image segment by segment, Amal said.
A good analogy, Amal explained, would be to think of a CNN acting like an astronomer analyzing an individual star with a telescope. A vision transformer model, by contrast, acts more like our eyes that can take in the expanse of the night sky all at once.
"A CNN is more local and the transformer gives a broader view," Amal said.
Clinical testing must validate the findings
Amal acknowledged that there were limits to the study. One of the biggest is that these models were only analyzed on one data set.
Further testing on multiple data sets is required to validate these findings. Amal said he and his team have begun working directly with clinicians to test these systems.
Amal said the team will validate the findings by working directly with pathologists, doctors who diagnose diseases by analyzing tissue and cell samples, and gastroenterologists, doctors who specialize in the digestive system.
"Pathologists help verify the AI's findings through tissue examination, while gastroenterologists assess how those findings can guide patient care," Amal said.
"Working with both groups in prospective studies allows us to evaluate these algorithms in real time, within clinical workflows, and determine how accurate, helpful, and useful they are in practice," he added.
Leon Toussaint, scientific manager at the National Foundation for Cancer Research, a U.S. nonprofit that supports cancer lab research, lauded Northeastern's findings.
"Research like this, which rigorously identifies where AI succeeds and where it fails in imbalanced real-world data, is exactly the kind of foundational work needed to bring reliable AI screening tools into clinical practice and prevent more cancers before they start," he said.
More information
Rajashekar Korutla et al., Transformer Architectures Improve Minority-Class Detection in Histopathological Classification of Colorectal Polyps, Lecture Notes in Computer Science (2026). DOI: 10.1007/978-3-032-35393-1_19
Who's behind this story?
Lisa Lock
BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021. Full profile →
Robert Egan
Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →
This story is republished courtesy of Northeastern Global News news.northeastern.edu.
Citation: AI is helping pathologists find harmful lesions the eye can't see (2026, October 20) retrieved 20 October 2026 from https://medicalxpress.com/news/2026-10-ai-pathologists-lesions-eye.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.
来源:Medical Xpress · medicalxpress.com