sFRC for assessing hallucinations in medical image restoration

A team of researchers from the U.S. Food and Drug Administration (FDA) and the National Institutes of Health (NIH) has published a new paper proposing sFRC (scanning Fourier Ring Correlation), a novel method for detecting hallucinations in AI-generated medical images. The work addresses a critical safety gap: while deep learning (DL) models like those for CT super-resolution or MRI restoration can produce visually appealing results, they often invent or 'hallucinate' anatomical features that don't exist in the original scan, posing a direct risk to patient diagnosis. Current methods lack robust, easy-to-use metrics to identify these fabrications, making sFRC a significant step toward trustworthy AI in clinical settings.

The sFRC technique works by performing Fourier Ring Correlation analysis over small patches and scanning across the AI output and a reference image to detect inconsistencies. The researchers validated sFRC on three undersampled medical imaging problems—CT super-resolution, CT sparse view, and MRI subsampled restoration—demonstrating its effectiveness in detecting hallucinated features. They also used it to quantify hallucination rates of DL methods on different data types and under increasing subsampling rates, providing a new way to characterize model robustness. Beyond DL, sFRC was shown to detect hallucinations in conventional and state-of-the-art unrolled restoration methods, making it a versatile new tool for the medical AI validation toolkit.