Discrete Diffusion Language Models for Interactive Radiology Report Drafting

Faster, Bidirectional Text Generation for Radiology Reports

Radiology reports are tedious to write—they're verbose, repetitive, and often need revision. This paper argues that diffusion language models (which build text by gradually refining noise into coherent tokens) can do the job better than the standard left-to-right autoregressive models that dominate medical AI. The authors demonstrate that a finetuned diffusion model not only matches the accuracy of its autoregressive peer but also unlocks a workflow advantage: radiologists can highlight fragments of a report and have the model intelligently fill in the gaps—something autoregressive models do poorly.

How It Works

Autoregressive models generate one token at a time, left to right, like typing. Diffusion models work differently: they start with a noisy "canvas" of random tokens and iteratively refine it across the entire text at once. This bidirectional denoising is the key. A radiologist can pin down, say, the clinical history and impression sections while leaving the findings section blank. The diffusion model then fills the findings with text consistent with the pinned sections—an operation called any-order infill. Autoregressive models can't do this naturally; they'd need awkward workarounds like masking or restarting generation.

The authors adapted DiffusionGemma-26B, a mixture-of-experts diffusion model, using LoRA (Low-Rank Adaptation), a lightweight finetuning technique that adapts only a small subset of parameters. They benchmarked it against Gemma-4-26B, an autoregressive baseline of the same size, on medical visual question answering datasets drawn from radiology. Both models were trained under identical conditions. An LLM judge scored responses in a way robust to length differences, since diffusion and autoregressive models often vary in verbosity.

The results: diffusion matched or exceeded autoregressive performance across all test datasets. The finetuned diffusion model ran with only 3.8 billion parameters active (the rest dormant in the mixture-of-experts layers) and remained competitive with large frontier vision-language models like GPT-4V. Crucially, decoding ran 3.5–4.4 times faster than the autoregressive baseline.

Why This Matters

Speed and accuracy are table stakes in medical AI, and the diffusion model delivers both. But the infill capability is the real win. Real radiology reports are messy: clinicians abbreviate, skip sections, or phrase findings differently across institutions. A radiologist who has dictated a rough report (or edited one from a template) can now lock key sections and let the model complete the rest intelligently. This mirrors how radiologists actually work—they iterate and refine—rather than forcing them into a linear generation workflow. The paper shows that diffusion isn't just a curiosity; it's a practical alternative to autoregressive dominance in medical contexts. The authors' insight is simple but overlooked: medical foundation models have ignored diffusion, but the bidirectional denoising process is well suited to the incremental, context-dependent nature of radiology documentation.