NGSE-Corr ranks imaging precision without a gold standard

On the same clinical image, three tumor-delineation methods can draw three different volumes. The actual tumor size is usually unavailable, leaving doctors and researchers without a clean yardstick. At Washington University in St. Louis, Abhinav Jha and his team have developed NGSE-Corr, a technique designed to rank the precision of medical imaging measurements without that missing gold standard.

The method tackles a subtle problem in repeated measurement. If several tools assess the same tumor, random fluctuations in their outputs may be correlated rather than independent. Jha modified an earlier mathematical formulation to account for that shared noise, then used numerical experiments to test whether NGSE-Corr could sort methods by precision.

The team next ran a virtual imaging trial involving three quantitative methods for measuring regional activity uptake in computer-generated patients with bone metastatic castrate-resistant prostate cancer treated with radium-223. Without knowing the ground truth—the actual value the tools were meant to measure—the technique ranked the methods correctly in 91% of trials conducted in groups of 50 virtual patients. It identified the most precise method in 95% of trials, with larger patient groups improving the results further.

What this could change: Researchers could use clinical data to compare new imaging tools, including AI-based systems, without first obtaining an expensive or time-consuming reference measurement. Physicians may gain another way to judge competing measurement methods, while regulators could use the approach when evaluating new medical imaging technologies. The technique has not yet been validated with clinical data: the reported results come from numerical experiments and computer-generated patients.

The project was led by Jha, an associate professor of biomedical engineering at Washington University's McKelvey School of Engineering and of radiology at WashU Medicine's Mallinckrodt Institute of Radiology. Yan Liu was the study's first author, with Daniel L. J. Thorek, Barry A. Siegel and Jingqin Luo as co-authors. The findings were published in IEEE Transactions on Medical Imaging.