Pitt researchers find mammogram AI can flag uncertainty without losing accuracy in initial tests

Breast cancer AI that gets most diagnoses right still needs to signal which answers deserve less trust. University of Pittsburgh researchers reported an initial step: one mammogram-analysis model estimated patient-level uncertainty without reducing overall diagnostic accuracy. Pittsburgh Supercomputing Center described the finding on October 15, following the team’s February 2026 presentation.

The first analysis covered several networks using uncertainty-weighted learning, or UWL. According to the center, one gave useful assessments of uncertainty that could cue doctors and patients when a diagnosis was less trustworthy. That result does not establish an improvement in patient care.

The team presented its results at the Society of Photo-Optical Instrumentation Engineers meeting in Vancouver in February 2026. The associated paper, identified in Medical Xpress’s October 21 coverage, is titled “Uncertainty-weighted learning to reduce diagnostic uncertainty, evaluate AI models, and augment precise clinical decisions.”

Teaching the model to account for doubt

The project involved PhD student Jiren Li, working with Pitt radiology professor Shandong Wu and research assistant professor Dooman Arefan. Their starting concern was the distinction between getting diagnoses right across a population and indicating how reliable a particular patient’s result might be.

UWL uses a deep neural network—an AI model trained on labeled examples—to analyze mammograms. The network is designed to produce two outputs: a diagnosis and an uncertainty score associated with that prediction.

That score also participates in training. Rather than serving only as information delivered alongside a finished prediction, uncertainty guides the network’s learning on uncertain examples. The method connects confidence in an answer with the process used to teach the model.

High-resolution images, substantial computing needs

Developing the networks used the center’s Bridges-2 supercomputer through an allocation from ACCESS, the National Science Foundation’s network of supercomputing sites. Wu described Bridges-2 as a longstanding resource for his lab’s model development, evaluation and medical-image storage.

Mammograms are large, high-resolution images. Training on them required graphics processing units, or GPUs, with substantial shared memory, plus storage for the imaging collection. The center describes that memory as a way to reduce slowdowns from repeatedly accessing data storage while processing training images.

Proposed next directions include combining medical images with other medical-record data and tailoring human–AI decision-making to measured model confidence. These are extensions the researchers want to pursue, rather than demonstrated outcomes of the initial analysis.