New AI Method Gauges Reliability in Breast Cancer Diagnosis by Measuring Uncertainty for Each Patient

University of Pittsburgh Tackles AI Reliability Crisis in Medical Imaging

Researchers at the University of Pittsburgh have developed a breakthrough approach to address a critical weakness in AI breast cancer diagnosis: the inability to gauge confidence in individual predictions1

1

2

2

. Led by Shandong Wu, a professor of radiology at the University of Pittsburgh School of Medicine, the team presented their findings at the Society of Photo-Optical Instrumentation Engineers (SPIE) annual meeting in February 2026 in Vancouver, Canada2

2

. The research addresses growing concerns about AI reliability in healthcare, where even rare diagnostic errors can have catastrophic consequences for patients.

Why Uncertainty-Weighted Learning Matters for Patient Safety

Traditional deep neural networks excel at providing accurate diagnoses across patient populations but struggle with a fundamental limitation: they cannot reliably indicate how confident they are about individual patient diagnoses1

1

3

3

. "I'm trying to develop effective AI models or methodology that can make AI model development easier, make breast cancer diagnosis more accurate, and make this model, make medical imaging AI models in general, more trustworthy and thus more clinically useful," Wu explained1

1

. This gap becomes critical in life-threatening diseases where clinicians need to know not just what the AI predicts, but how much they should trust that prediction for the specific patient in front of them.

Uncertainty-Weighted Learning Transforms Deep Neural Network Architecture

Jiren Li, a PhD student working with Wu and research assistant professor Dooman Arefan, pioneered the uncertainty-weighted learning (UWL) approach that fundamentally redesigns how deep neural networks process mammogram images2

2

3

3

. The UWL architecture generates both a diagnosis and an associated uncertainty score when analyzing breast mammogram images1

1

. This confidence score becomes integral to the learning process itself, guiding the model to handle uncertain samples more effectively during training. The innovation addresses the reality that while AI can be correct most of the time, giving wrong answers with equal confidence undermines clinical utility and patient trust.

Bridges-2 Supercomputer Enables Computational Breakthrough

Implementing uncertainty-weighted learning for AI breast cancer diagnosis required substantial computational resources that only advanced supercomputing infrastructure could provide. The team utilized the Bridges-2 supercomputer at Pittsburgh Supercomputing Center through an allocation from ACCESS, the NSF's network of supercomputing sites2

2

3

3

. Bridges-2 features 47 GPU nodes containing 384 late-model GPUs, with each node providing between 128 and 2,000 gigabytes of RAM—dramatically exceeding the 64 GB found in high-end laptops1

1

3

3

. Processing high-resolution mammogram images demands this level of GPU memory to prevent slowdowns from repeatedly accessing data storage during computations.

Medical Imaging AI Requires Specialized 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," Wu noted1

1

3

3

. The computational challenges stem from the nature of breast mammograms themselves—large, high-resolution images that require graphics processing units with substantial shared memory access. PSC's Ocean active project data storage system provided the necessary capacity to host complete libraries of imaging data for AI model training and testing, enabling the team to develop and evaluate multiple AI models over several years.

Broader Implications for Trustworthy AI in Healthcare

The development arrives at a pivotal moment for AI reliability across industries. While AI powers beneficial applications from online shopping to fraud prevention, high-profile failures—including AI-generated misinformation, autonomous vehicles failing to recognize pedestrians, and biased algorithmic behavior—have raised legitimate concerns about trustworthiness2

2

3

3

. In medical imaging, where AI assists radiologists in identifying breast cancer earlier and predicting treatment needs, the stakes are particularly high. The uncertainty-weighted learning approach offers a path forward: rather than simply improving average accuracy, it provides clinicians with the context they need to make informed decisions about individual patients, potentially transforming how AI tools achieve clinical utility in healthcare settings where lives depend on getting the diagnosis right.