New AI System Weighs Uncertainty in Breast Cancer Diagnosis Using Deep Neural Networks

Addressing AI Reliability in Medical Diagnosis

Researchers at the University of Pittsburgh have developed a groundbreaking approach to AI breast cancer diagnosis that directly confronts a critical weakness in current medical AI systems. Led by Shandong Wu, a professor of radiology at the University of Pittsburgh School of Medicine, the team created a deep neural network that doesn't just provide diagnoses—it also calculates how confident the system is in each prediction for individual patients.1

1

The innovation addresses a fundamental problem with existing AI systems: while they may be accurate most of the time across populations, they struggle to assess reliability for specific cases. When dealing with life-threatening diseases like breast cancer, even occasional errors can have catastrophic consequences for affected patients. Wu explains his goal is to "develop effective AI models or methodology that can make AI model development easier, make breast cancer diagnosis more accurate, [and] make medical imaging AI models in general, more trustworthy and thus more clinically useful."2

2

Uncertainty-Weighted Learning Changes the Game

The solution came through what the team calls uncertainty-weighted learning, or UWL. This approach fundamentally redesigns how the deep neural network processes breast mammogram images. Instead of simply outputting a diagnosis, the system simultaneously generates an "uncertainty score" that quantifies the reliability of each prediction.1

1

Jiren Li, a PhD student working with Wu and research assistant professor Dooman Arefan, spearheaded the development of this novel architecture. The uncertainty score isn't just an add-on—it becomes integral to the learning process itself. During training, the system uses these scores to guide how it handles uncertain samples, creating trustworthy AI models that can distinguish between high-confidence and low-confidence predictions.2

2

This represents a shift from population-level accuracy to individual patient assessment, providing a confidence score for individual patients rather than relying on aggregate statistics. For radiologists using AI in medical imaging, this means knowing when to trust the system's output and when additional human review is essential.

Computational Challenges Required Supercomputing Power

Implementing uncertainty-weighted learning posed significant computational challenges that required the Bridges-2 supercomputer at Pittsburgh Supercomputing Center (PSC). The system needed to process high-resolution mammogram images, which are large files requiring substantial computing resources. Graphics-processing units (GPUs) handle this type of work best, but the team needed GPUs with access to large, shared memory to avoid slowdowns from repeatedly accessing data storage.1

1

The Bridges-2 supercomputer delivered exactly what the researchers needed: 47 GPU nodes containing 384 late-model GPUs, with each node equipped with between 128 and 2,000 gigabytes of RAM per node—vastly exceeding the 64 GB found in high-end laptops. The team accessed Bridges-2 through an allocation from ACCESS, the NSF's network of supercomputing sites.2

2

Wu emphasized the critical role of the infrastructure: "Bridges-2 is critical for the success of these studies, because as you can imagine, this is all using large-scale, high-resolution, heterogeneous radiology imaging data, and in order to build this AI model, we need GPUs, memory, and larger storage to make this possible."1

1

The GPU memory requirements and massive data storage capacity of PSC's Ocean system enabled the team to store complete libraries of imaging data for AI model training and testing.

Why This Matters for Medical AI's Future

The development comes at a crucial moment for AI in healthcare. While AI systems have demonstrated impressive capabilities in medical imaging and diagnosis, concerns about reliability persist across the broader AI field. From autonomous vehicles failing to recognize pedestrians to AI systems producing fabricated information with unwarranted confidence, the technology's limitations have become increasingly apparent.2

2

For breast cancer diagnosis specifically, the stakes are exceptionally high. Early detection significantly improves treatment outcomes, but misdiagnoses—whether false positives or false negatives—carry serious consequences. A reliable AI for breast cancer diagnosis must not only be accurate on average but also signal when its predictions are less certain, allowing clinicians to make informed decisions about when to seek additional confirmation.

The University of Pittsburgh team's work on clinically useful AI systems could establish a template for developing trustworthy medical AI across other diagnostic applications. By building uncertainty quantification into the architecture from the ground up rather than adding it as an afterthought, the approach creates a foundation for AI systems that acknowledge their own limitations—a critical characteristic for any technology deployed in clinical settings where human lives depend on accurate information.