Researchers have built an AI model that generates realistic 3D CT scans straight from the text of a radiology report.
Earlier text-to-CT systems paired their 3D image generators with encoders trained only on language or on flat 2D images, so the model reading the report had never actually seen a volumetric scan. The new method fixes that by training a 3D-CLIP encoder on structured hard negatives at the text level, sharpening its sense of what a finding should look like in three dimensions without needing more GPU memory to do it. That encoder conditions an end-to-end latent diffusion model that builds the CT volume directly in 3D space, skipping the slice-by-slice super-resolution step that causes visible seams in older pipelines. Tested on the CT-RATE benchmark across 18 pathological conditions, the method beats prior approaches on both image fidelity and factual correctness, while using less inference time and GPU memory; the code is on GitHub.
Medical imaging AI is chronically short on labeled 3D data, and a text-conditioned generator that actually respects anatomy could help simulate rare conditions or pad out thin training sets. The efficiency claim matters just as much as the accuracy claim, since volumetric diffusion models are normally too memory-hungry for anyone outside a well-funded lab to run.
It is still a benchmark paper, not a bedside tool, and a scan that looks anatomically correct is not the same as one a radiologist should trust with a diagnosis.