AI quality control system validates medical image synthesis for clinical safety

In radiotherapy clinics and treatment planning centers, doctors increasingly rely on synthetic medical images—pictures generated by artificial intelligence rather than captured directly from patients. These images fill gaps: reconstructing a missing scan modality, adapting a treatment plan to anatomical changes, or synthesizing a CT scan from an MRI when the original isn't available. The problem is invisible. A generative model can produce an image that looks plausible to the human eye but contains subtle artifacts—distortions in tissue boundaries, misaligned intensity contrasts, anatomical inconsistencies—that a clinician might miss. In radiotherapy, where millimeter-level precision determines whether radiation hits the tumor or damages healthy tissue, a flawed synthetic image could alter a treatment plan in ways that harm the patient. No one would know until the damage was done.

A team of researchers set out to solve this by building an automated quality control system that could reliably flag when a synthetic medical image was safe to use and when it wasn't. They started with a generative model called SynDiff, an adversarial diffusion-based framework, and tested it on four different image synthesis tasks: three that translated between different types of MRI scans, and one that synthesized CT images from cone-beam CT scans. The goal was straightforward but demanding: create a machine learning system that could predict whether a radiologist would accept a synthetic image as clinically usable.

To train this system, the researchers needed ground truth. They collected visual quality ratings from thirteen independent raters, each scoring hundreds of synthetic images on a six-point scale using a standardized protocol and specialized viewing software. The raters worked blind, randomized, and independently—no collaboration, no bias toward any particular image. This produced a consensus judgment: a distribution of human opinion about which images were acceptable and which were not.

Then came the modeling work. The team computed eighteen different automated quality metrics for every synthetic image—ten that compared the synthetic image to an original reference image, and eight that assessed quality without any reference at all. These metrics measured things like structural similarity, anatomical boundary preservation, and the contrast relationships between different tissue types. They fed all these numbers into Auto-Sklearn, an ensemble machine learning system, and trained it to predict what the thirteen human raters would say about each image.

The results were encouraging. The reference-based models—those using metrics that compared synthetic images to originals—achieved an R² of 0.75, meaning they explained 75 percent of the variance in human ratings and typically predicted ratings within ±0.5 points on the six-point scale. The no-reference models, which had to assess quality without seeing the original image, performed more modestly at R² of 0.59, but remained unbiased and informative. This matters because in real clinical practice, you often don't have a reference image to compare against. A synthetic image might be the only version available.

Explainability analysis revealed which metrics mattered most. Measures of structural fidelity—how well anatomical boundaries were preserved, how accurately tissue contrasts were reproduced—emerged as the strongest predictors of human acceptability. The system wasn't just making predictions; it could explain why it thought an image was good or problematic.

What the researchers demonstrated is that ensemble regression models can serve as a transparent, scalable quality gate for generative medical imaging. Before a synthetic image enters a treatment planning workflow, it passes through this automated checkpoint. The system flags images that fall below clinical acceptability thresholds, preventing flawed synthetics from reaching the radiologist's screen. For adaptive radiotherapy—where treatment plans are adjusted mid-course based on new imaging—or for missing-modality reconstruction, where one imaging modality is unavailable, this kind of automated validation could be the difference between safe deployment and undetected harm. The work suggests a path forward: generative AI in medicine doesn't have to be a black box. With the right quality control infrastructure, it can be as transparent and clinically accountable as any other tool in the treatment room.