How trustworthy is AI breast cancer detection? New method gauges uncertainty for individual patients
While AI offers the ability to improve human experts' ability to diagnose and treat diseases like breast cancer, well-publicized errors across the broader AI field are cause for concern. A team from the University of Pittsburgh has used PSC's flagship Bridges-2 supercomputer to design a deep neural network (DNN) that takes uncertainty into account in diagnosing breast scan images.
A new method from the University of Pittsburgh aims to improve the trustworthiness of AI breast cancer detection by incorporating uncertainty assessment for individual patients. This approach, developed using the university's Bridges-2 supercomputer, provides a gauge of how accurate the AI's predictions are for each patient, rather than just the average across a population.
The researchers, led by Shandong Wu, a radiology professor, and Jiren Li, a Ph.D. student, hope this method will make medical imaging AI more reliable and clinically useful. The system, which uses uncertainty-weighted learning (UWL), provides both a diagnosis and an associated uncertainty score when analyzing breast mammogram images.
This uncertainty score is an integral part of the learning process, guiding the model's learning to handle uncertain samples more effectively. The Pitt team faced significant computational challenges in developing this method, requiring access to large, shared memory and extensive data storage to process high-resolution imaging data.
Bridges-2, a supercomputer with 47 GPU nodes and 15 petabytes of storage, was crucial for these studies. The results of their initial analysis using UWL-based DNNs showed promising outcomes, with one DNN providing accurate estimates of uncertainty without compromising overall accuracy. This method has broader applications beyond breast cancer diagnosis, potentially making many critical AI applications more trustworthy and reliable.
Written by urgent.news from Medical Xpress's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
