Breast cancer AI now flags its own doubts, in a lab study

Ask an AI model about a mammogram and it returns a diagnosis. It does not say how much to trust that particular answer. A team from the University of Pittsburgh has built a model that does: a deep neural network, or DNN, that gives each patient a diagnosis together with an uncertainty score. The work was presented at the annual meeting of the Society of Photo-Optical Instrumentation Engineers (SPIE) in February 2026, in Vancouver, Canada.

DNNs learn from training sets in which humans have labeled the right answers. They become very good at predicting the likely answer across a population of patients, sometimes almost all of the time. They are much weaker at telling how reliable a diagnosis is for one person. For a disease that threatens life, the team points out, even a few wrong diagnoses are catastrophic for the patients they affect. The team’s method, which it calls uncertainty-weighted learning (UWL), feeds the uncertainty score into training itself, so the model learns to handle uncertain cases more carefully.

The obstacle was computing. Mammograms are large, high-resolution images, and graphics processing units (GPUs), the chips DNNs run best on, need large shared memory to process each one without constant trips to storage. Jiren Li, a PhD student, obtained time on Bridges-2, the supercomputer of PSC, through ACCESS, the network of supercomputing sites run by the US National Science Foundation (NSF). Bridges-2 offered 47 GPU nodes with 384 late-model GPUs, and each node has between 128 and 2,000 gigabytes of RAM, against 64 GB in a high-end laptop. Its Ocean storage system holds 15 petabytes of usable data.

The first analysis, which compared several UWL-based DNNs, was encouraging. One of them gave good estimates of its own uncertainty, signaling to doctors and patients when a diagnosis may be less reliable for a specific patient. It did this without losing accuracy across the population as a whole. The reporting gives no accuracy figures for the model, and the team describes this as a first analysis, not a clinical validation.

For radiologists, the gain would be a second signal next to each answer, showing which cases deserve a closer read. For patients, it would be a warning when a diagnosis is less reliable for them personally. The team says the same approach could apply to other critical uses of AI, beyond breast cancer. It gives no timeline for clinical use, and the model is still at the research stage.