AI wrapper flags uncertainty in lung cancer slide classification
Researchers say TRUECAM can screen uncertain pathology images and defer difficult cancer subtype calls to human specialists.
By Tom Brennan · Health & Medicine Correspondent
3 min read
Researchers at Vanderbilt Health and institutions in Hong Kong have developed an artificial intelligence framework meant to make digital pathology systems less likely to issue confident but unreliable cancer subtype calls. The work matters because medical AI models can misclassify unfamiliar or low-quality inputs when they are pushed beyond the data they were trained on.
The framework, called TRUECAM, was described in a peer-reviewed paper in Nature Biomedical Engineering. Vanderbilt University Medical Center said the system is designed as a wrapper, meaning it sits around existing pathology AI tools to manage how they process images and report results.
The researchers demonstrated TRUECAM mainly on non-small cell lung cancer subtyping from whole-slide pathology images, according to Vanderbilt University Medical Center. The group tested it with a widely used AI architecture for that task and with four newer digital pathology foundation models.
Filtering weak evidence before classification
Medical AI systems often lack a reliable way to measure their own uncertainty, according to the researchers. Vanderbilt University Medical Center said neural networks may give confident answers even when an image differs from their training data, much as an animal classifier trained on one region’s species might force an unfamiliar animal into a known category.
TRUECAM was built to detect inputs that fall outside a model’s intended scope and to remove parts of slides that are unlikely to help diagnosis, the paper reports. Those regions can include normal tissue or poorly stained material that may interfere with a slide-level prediction.
According to the study, that filtering lets TRUECAM focus on diagnostically useful image patches and provide user-specified accuracy targets for subtype classification. Chao Yan, a research instructor in biomedical informatics at Vanderbilt and one of the paper’s three lead authors, said the framework often concentrated on the same areas that pathologists would consider relevant.
Tests across hospitals and cancer types
The team evaluated TRUECAM on non-small cell lung cancer whole-slide images from two geographically diverse cancer research consortia, Vanderbilt University Medical Center said. The researchers also used a constructed set of clinically meaningful out-of-scope images and a sequence of real-world images from Queen Mary Hospital in Hong Kong.
Testing extended beyond lung cancer to tissue images from several organs, including breast, brain and kidney, according to the paper. The authors reported that TRUECAM generalized to datasets outside lung cancer while maintaining its uncertainty-screening functions.
The study authors said TRUECAM outperformed existing approaches to uncertainty quantification in digital pathology AI in accuracy and efficiency. Vanderbilt University Medical Center said the method worked relatively quickly, did not add substantial costs and reliably identified images that were outside a model’s scope.
The paper also reported that TRUECAM could abstain from classifying difficult cases, leaving those slides for pathologists to review. The authors said its error rates met prespecified targets and that the framework improved fairness across sex and race in the tested datasets.
Bradley Malin, a Vanderbilt professor of biomedical informatics, biostatistics and computer science and one of the paper’s corresponding authors, said trustworthy medical AI requires systems that can recognize when patient data, tissue preparation methods or image artifacts make a model’s answer unsafe. He said TRUECAM offers a practical way to address those failure modes across digital pathology tools.
This story draws on original reporting from Medical Xpress.