RADiFF: Anatomy-Block Masked Diffusion for Structured Radiology Report Generation

RADiFF generates the structured Findings section of a chest X-ray report, one anatomical section at a time. Within each section the text is produced by masked discrete diffusion: all positions start masked and are filled in parallel over 3 denoising steps. A single shared denoiser writes all eight SRRG-Findings categories:

Lungs and Airways · Cardiovascular · Pleura · Musculoskeletal and Chest Wall · Tubes, Catheters, and Support Devices · Hila and Mediastinum · Other · Abdominal

Each section is conditioned on the image and on the sections already written. A per-category emission threshold decides whether a section is written at all.

This repository contains the image-only model evaluated in the paper.

Model details

Image encoder	microsoft/rad-dino, frozen (downloaded separately, pinned revision)
Text denoiser	6-layer transformer decoder, d=768, 8 heads, cross-attention to image patches
Input	1-2 views per study (frontal first), resized and centre-cropped to 518×518
Output	SRRG-Findings structured report (Category: headers with - bullets)
Tokenizer	BPE, 8,192 tokens, trained on SRRG-Findings train reports
Decoding	3 steps per section; thresholds in config.json
Training data	StanfordAIMI/srrg_findings train split (181,874 studies, MIMIC-CXR + CheXpert Plus)

model.safetensors holds every weight except the frozen RAD-DINO backbone, which is loaded from its own repository at revision 110cbc18d5133582e320b43d53bf5c44e410c936.

Usage

pip install -r requirements.txt
python inference.py --images frontal.jpg lateral.jpg
python inference.py --studies studies.txt --output reports.jsonl   # one study per line, comma-separated

from inference import load_radiff, generate_reports

model, tok, cfg = load_radiff(".", device="cuda")
reports = generate_reports(model, tok, cfg, [["frontal.jpg", "lateral.jpg"], ["other_frontal.jpg"]])
print(reports[0])

Example output format:

Lungs and Airways:
- No focal consolidation.

Cardiovascular:
- Normal cardiomediastinal silhouette.

Pleura:
- No pleural effusion or pneumothorax.

The released package reproduces the paper's Test and Test Reviewed predictions exactly.

Results

SRRG-Findings, image-only models. Scores use StructEval (aligned=False, lower-cased, leaves taxonomy). R-L = ROUGE-L, BERT = BERTScore, RG = F1-RadGraph, SRR = F1-SRR-BERT.

Split	Model	BLEU	R-L	BERT	RG	SRR	Cat P	Cat R	Cat F1
Test	CheXpert-Plus	1.72	18.61	31.12	14.42	39.28	72.73	86.00	76.98
Test	CheXagent-2-3b	1.76	19.54	31.38	15.33	40.16	77.65	81.17	77.46
Test	MAIRA-2	1.52	19.22	31.08	15.14	39.48	73.19	80.63	75.11
Test	RADiFF	1.40	20.55	32.23	16.16	42.78	77.87	84.17	79.33
Test Reviewed	CheXpert-Plus	2.13	18.61	30.93	14.71	39.02	71.44	88.46	76.91
Test Reviewed	CheXagent-2-3b	2.41	19.71	32.17	15.80	41.10	75.21	84.65	77.17
Test Reviewed	MAIRA-2	0.92	18.73	31.13	15.14	40.32	70.93	83.95	75.15
Test Reviewed	RADiFF	1.17	20.48	32.53	16.54	45.10	77.53	89.99	81.33

Intended use and limitations

• Research only. RADiFF is not a medical device and must not be used for diagnosis, treatment or any clinical decision.
• Trained only on MIMIC-CXR and CheXpert Plus studies (two US academic centres); performance on other populations, scanners or acquisition protocols is unknown.
• Because many positions are filled in parallel, generated sentences can contain repeated words or incomplete phrases (e.g. "No pleural effusion effusion."). Reports should be read as structured findings, not as finished prose.
• Like any report generator, it can omit findings or state findings that are not present.

License

RADiFF Research-Only, Non-Commercial License.

Citation

@misc{alas2026radiff,
  title     = {{RADiFF}: Anatomy-Block Masked Diffusion for Structured Radiology Report Generation},
  author    = {Alas, Mohamed and Obrador-Reina, Miquel and Marhuenda, Luis-Jesus and Albiol, Alberto and Paredes, Roberto},
  year      = {2026}
}