New TRUECAM wrapper makes medical artificial intelligence more trustworthy

Medical artificial intelligence (AI) faces a fundamental challenge: uncertainty quantification. Artificial neural networks are largely unaware of the limits of their training data and can become overconfident when confronted with unfamiliar inputs. Suppose you train a neural network to distinguish among African mammals. If you then present it with an image of a South American jaguar - an animal it has never encountered - the model cannot say "unknown." Instead, it may confidently declare the jaguar to be a leopard.

Reported June 23 in Nature Biomedical Engineering, researchers at Vanderbilt Health and centers in Hong Kong have created a versatile uncertainty-aware AI framework broadly adaptable as a wrapper for digital pathology AI systems. (An AI wrapper acts as an interface layer that customizes, formats and automates how users interact with the underlying intelligence.) They demonstrate their wrapper, called TRUECAM, primarily with reference to non-small cell lung cancer (NSCLC) subtyping using whole-slide images.

TRUECAM is designed not only for identification of out-of-scope inputs, but also to filter out noninformative regions - normal or poorly stained tissue, for example - that could distort slide-level inference. According to the paper, these complementary capacities allow TRUECAM to provide customizable accuracy guarantees for cancer subtype classifications.

The team tested TRUECAM as a wrapper for a widely used AI architecture for NSCLC subtyping and four newer, more generalized digital pathology AI foundation models. Testing involved NSCLC whole-slide images from two geographically diverse cancer research consortia, 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 also extended to cancer tissue spanning multiple organs, such as breast, brain and kidney.

Compared to other solutions for trustworthy digital pathology AI, the authors say that TRUECAM performs not only with greater accuracy, but also relatively rapidly and efficiently, without adding substantial costs.

Achieving trustworthy AI in the medical domain is requisite for realizing the potential of this transformative technology. It's not only that a patient's clinical profile can fall out of scope of your model's training data, but other sources of variation, such as your institution's method of collecting and staining specimens, or artifacts and irregularities that tend to arise in tissue preparation, can also prompt your model to arrive confidently at a mistaken conclusion. TRUECAM provides a thoroughgoing, versatile and efficient solution to these potentially unsafe shortcomings."

Bradley Malin, PhD, one of the paper's three corresponding authors, Professor of Biomedical Informatics, Biostatistics and Computer Science, and holder of the Accenture Chair

The team reports that TRUECAM proved more accurate and efficient than existing approaches to digital pathology AI uncertainty quantification, reliably detected out-of-scope inputs, abstained from classifying challenging inputs (allowing deferral to pathologists), delivered error rates that reliably met prespecified accuracy targets, improved fairness across sex and race, and proved generalizable to datasets beyond lung cancer.

Chao Yan, PhD, MS, Research Instructor in Biomedical Informatics, is among the paper's three lead authors.

"Perhaps our most striking finding," Yan said, "was that, with ambiguous patches and normal regions often found to dominate a pathology slide, TRUECAM's targeted elimination of this noise, and its resulting focus on sometimes comparatively small patches in an image, allows it to proceed efficiently to accurate and fairer cancer subtype classification, with the model focusing on the same regions pathologists identify as diagnostically relevant. This goes beyond current approaches, and the practical implications appear to have broad import."

Chief authors from Hong Kong include, from Hong Kong Polytechnic University, lead author and corresponding author Xiaoge Zhang, PhD, who earned his doctorate from Vanderbilt University in 2019, and lead author Tao Wang, a PhD student, and, from the University of Hong Kong, corresponding author Maximus C.F. Yeung, MBBS, MSc. Also on the paper from Vanderbilt Health is Fedaa Najdawi, MBBS, Assistant Professor of Pathology, Microbiology and Immunology. The study was funded in part by the National Institutes of Health under award K99LM014428.