DRRG: A Discrete Diffusion Framework for Radiology Report Generation

Shaoyang Zhoua; Yingshu Li; Yunyi Liu; Lijun Pu; Lingqiao Liu; Lei Wang; Luping Zhou

Source-linked AI summary

DRRG: A Discrete Diffusion Framework for Radiology Report Generation

Shaoyang Zhoua, Yingshu Li, Yunyi Liu, Lijun Pu, Lingqiao Liu, Lei Wang, Luping Zhou

arXiv:2608.24105v1Aug 25, 2026cs.CV

TL;DR

Radiology report generation commonly uses left-to-right decoding that cannot revise earlier content. DRRG instead uses discrete diffusion for iterative refinement and demonstrated effectiveness across evaluation metrics on MIMIC-CXR and CheXpert Plus.

Problem
Left-to-right autoregressive reporting can amplify early errors and does not match radiology’s iterative drafting and revision workflow.
Method
DRRG formulates report generation as iterative masked-token denoising with clinically focused masking and image-derived concept conditioning.
Results
DRRG demonstrated effectiveness across evaluation metrics on MIMIC-CXR and CheXpert Plus, with RadGraph-F1 of 0.281, 35.1% above the second-best method.
Takeaways & Limitations
Discrete diffusion offers an effective alternative to autoregressive radiology report generation by enabling bidirectional refinement aligned with revisable reporting.
Takeaways & Limitations
The study was limited to secondary analysis of two publicly available de-identified chest radiography datasets without recruiting new participants.

Abstract

from arXiv · show

Purpose: Automatic radiology report generation (RRG) has been widely explored to improve reporting accuracy and reduce radiologists' workload. Most existing methods rely on autoregressive (AR) frameworks that generate reports token by token and cannot revise earlier content, making them prone to error propagation and inconsistent with the iterative refinement process of radiological reporting. In contrast, discrete diffusion large language models (DLLMs) generate text through iterative denoising, naturally enabling report refinement. However, DLLMs have not been extensively investigated for RRG. In this study, we developed and evaluated a discrete diffusion framework for RRG that enables iterative refinement rather than conventional left-to-right autoregressive decoding. Materials and methods: We developed DRRG, a DLLM-based framework that formulates RRG as iterative masked-token denoising. DRRG incorporates a clinical-entities-aware complementary mask to improve token supervision coverage and emphasize clinically important entities, together with a concept-conditioning module that injects image-derived clinical concepts into visual representations. DRRG was trained and evaluated on MIMIC-CXR and CheXpert Plus. Results: On MIMIC-CXR, DRRG achieved BLEU-4 of 0.210, CheXpert-F1 of 0.549, RadGraph-F1 of 0.281, GREEN of 0.360, and RaTEScore of 0.604, outperforming the compared methods on most reported metrics, despite employing a substantially smaller LLM decoder. On CheXpert Plus, DRRG achieved the highest BLEU-4 (0.119) and CheXpert-F1 (0.347) among the compared methods. Conclusion: Discrete diffusion provides an effective alternative to autoregressive radiology report generation by enabling iterative, bidirectional report refinement. Incorporating clinically focused masking and image-derived concept conditioning improves report quality and clinical consistency.

1. Introduction

Radiology report generation is safety-critical, yet predominantly autoregressive methods can amplify early errors and offer limited revisability. DRRG addresses this gap with iterative masked-token denoising, clinically focused masking, and image-derived concept conditioning.

Motivation: Radiology reports require semantic consistency and logical correctness because incorrect negation or contradictory findings can lead to inappropriate patient management.
Limitations of Existing Methods: Autoregressive decoding fixes reports sequentially from left to right, allowing early errors to enter later context and limiting revision.This paradigm is also misaligned with the workflow in which reports are drafted, reviewed, and iteratively refined.
DRRG Framework: DRRG formulates radiology report generation as iterative masked-token denoising, enabling non-causal refinement with bidirectional report context.The framework is designed to better align generation with the revisable nature of radiology reporting.
Key Components: DRRG introduces a clinical-entities-aware complementary mask and a concept-conditioning module to emphasize clinically important entities and incorporate image-derived clinical information.The mask increases supervision coverage, while concept conditioning strengthens guidance during denoising.
Research Gap: Iterative generation paradigms that reflect radiology’s reporting process have received relatively limited attention despite advances in encoder–decoder, transformer, and LLM-based methods.

2. Materials and methods

DRRG is a multimodal discrete-diffusion framework for radiology report generation, evaluated on MIMIC-CXR18 and CheXpert Plus. It combines concept-conditioned visual representations, clinical-entities-aware complementary masking, and iterative denoising.

Model architecture: The model uses SigLIP as a vision encoder and augments visual features with 14 predicted clinical observations through concept embeddings and fusion.Mean-pooled visual features produce multi-label concept scores, whose probability-weighted embeddings are fused into visual tokens before projection into the language model’s hidden space.
Training objective: DRRG trains a discrete diffusion language model by replacing target-report tokens with [MASK] according to a sampled mask ratio and recovering the clean report from conditioned representations.The input combines concept-conditioned visual features, clinical context, and the ground-truth report; loss is computed on masked response tokens.
Masking strategy: Clinical-entities-aware complementary masking increases token-level supervision coverage while assigning higher masking probabilities to clinically important entity tokens.Complementary views preferentially mask tokens preserved in the other view, and RadGraph identifies clinical entity spans for enhanced masking.
Inference procedure: Inference starts from a fully masked report and iteratively denoises it, committing high-confidence tokens while re-masking low-confidence positions for later refinement.At each timestep, predictions are made simultaneously for all masked positions, and a K-step schedule determines how many newly committed tokens are retained.

3. Results

DRRG was evaluated across lexical, clinical, and LLM-based metrics on MIMIC-CXR and CheXpert Plus, outperforming compared methods on most reported measures. Qualitative analyses further examined iterative refinement and error correction through denoising.

Evaluation metrics: Evaluation covered lexical BLEU and ROUGE-L, clinical CheXpert and RadGraph-F1, and LLM-based GREEN and RaTEScore metrics.GREEN identifies clinically meaningful errors, whereas RaTEScore measures clinical entity correspondence between generated and reference reports.
Ablation and hyperparameters: Ablation and hyperparameter analyses evaluated the diffusion framework, clinical-entities-aware complementary mask, concept-conditioning module, and their effects on clinical and LLM-based metrics.Analyses also varied the clinical-entity masking factor, concept-loss weight, and initial generation length.
Qualitative analysis: Qualitative case studies compared AR and DLLM error patterns and visualized iterative refinement, showing additional denoising steps correcting incorrect descriptions.Figure 4 distinguishes incorrect, reference, and correct content with red, blue, and green highlights.

4. Discussion

DRRG outperformed competing radiology report generation methods across textual, clinical, and LLM-based metrics while using a smaller decoder. Ablations and case studies showed benefits from clinically focused masking, concept conditioning, and iterative denoising, with a quality–efficiency trade-off governed by generation settings.

NLG Metrics: DRRG achieved the best MIMIC-CXR BLEU-2, BLEU-3, BLEU-4, and ROUGE-L scores, while using a substantially smaller 0.6B text decoder than listed LLM-based competitors.Its scores were 0.332, 0.259, 0.210, and 0.343, respectively; BLEU-1 was second-best at 0.452.
Clinical Metrics: 0.558 precision and 0.549 F1 on MIMIC-CXR surpassed DART by 0.025 and 0.029, while RadGraph-F1 reached 0.281, 35.1% above MultiP-R2Gen.On CheXpert Plus, DRRG achieved the best CheXpert recall and F1, outperforming MambaXray-VL-L43 by 0.023 and 0.012.
LLM-based Metrics: 0.360 GREEN and 0.604 RaTEScore were the best LLM-based evaluation results, indicating stronger alignment with clinically grounded assessment criteria.These metrics further supported DRRG’s report quality beyond textual similarity and clinical-label evaluation.
Ablation Study: The DLLM-based model consistently outperformed the matched AR baseline across all reported NLG metrics, and adding clinical-entities-aware masking improved average NLG performance by 0.0024.The matched models used the same SigLIP2 vision encoder and Qwen3-0.6B text decoder; concept conditioning produced larger gains across NLG and clinical metrics.
Hyperparameter Analysis: Increasing the clinical-entity masking factor from 1.0 to 1.2 improved BLEU-4 from 0.205 to 0.210, ROUGE-L from 0.338 to 0.343, and RadGraph F1 from 0.278 to 0.281, whereas further increases reduced performance.Increasing concept-loss weight from 0.25 to 0.5 improved BLEU-4, ROUGE-L, CheXpert-F1, and RadGraph-F1 by 0.025, 0.014, 0.060, and 0.033.
Hyperparameter Analysis: 72 denoising steps improved all NLG metrics over 32 steps but increased inference time from 0.97s to 2.13s per case; 64 steps achieved higher scores at comparable runtime, while 32 steps used only 53.9%DRRG still outperformed the corresponding AR baseline with only 32 denoising steps.
Iterative Refinement: Iterative denoising allowed low-confidence tokens to be re-masked and revised, reducing error propagation compared with autoregressive generation and progressively correcting false findings.The case study reports noisy intermediate predictions at Step 24, greater coherence by Step 48, and further correction at Step 72.

5. Conclusion

DRRG is a discrete diffusion framework for radiology report generation that uses iterative denoising to enable bidirectional refinement. Its clinical-entities-aware complementary mask and concept-conditioning module improve data efficiency, clinical-entity awareness, and clinically guided denoising.

Framework: DRRG formulates radiology report generation as iterative denoising rather than autoregressive decoding, enabling bidirectional refinement.This formulation better aligns with the revisable nature of radiology reporting.
Clinical guidance: The clinical-entities-aware complementary mask improves data efficiency and enhances clinical-entity awareness.
Clinical guidance: The concept-conditioning module guides the denoising process with clinical concepts.It incorporates image-derived clinical concepts into visual representations.

CRediT authorship contribution statement

The authors contributed across conceptualization, methodology, software, data curation, validation, investigation, analysis, visualization, and manuscript writing. Shaoyang Zhou led the broadest set of listed contributions, while other authors contributed to methodological, validation, investigative, curation, and writing activities.

Author contributions: Shaoyang Zhou contributed to conceptualization, methodology, software, data curation, formal analysis, validation, visualization, and both original-draft and review-and-editing writing.These contributions are listed explicitly in the authorship statement.
Author contributions: Yingshu Li contributed to conceptualization, methodology, software, data curation, validation, and writing–review and editing.The statement attributes these roles directly to Yingshu Li.
Author contributions: Yunyi Liu, Lijun Pu, Lingqiao Liu, and Luping Zhou contributed across data curation, validation, investigation, methodology, conceptualization, and writing–review and editing as listed.The supplied statement identifies Yunyi Liu, Lijun Pu, Lingqiao Liu, and Luping Zhou with overlapping contributions in these areas.

Ethics statement

The study was a secondary analysis of two publicly available, de-identified chest radiography datasets and required no additional individual informed consent.

Ethics statement: The study analyzed publicly available, de-identified MIMIC-CXR and CheXpert Plus datasets without recruiting participants or accessing identifiable personal information.Data use followed the corresponding data-use requirements and institutional access procedures.

Declaration of competing interests

The authors declare no competing financial interests or personal relationships that could have influenced this work; Luping Zhou’s editorial-board role was excluded from the peer-review process.

Luping Zhou is on Meta-Radiology’s Editorial Board but did not participate in peer review, which was independently handled by another board member.All authors declared no competing financial interests or personal relationships that could have influenced the reported work.