RadHiera

[MICCAI2026] RadHiera: Semantic Hierarchical Reinforcement Learning for Medical Report Generation

RadHiera improves radiology report generation with report-level semantic rewards, impression-aware optimization, and Findings-Impression consistency alignment.

Released checkpoints

Dataset	Checkpoint	n	BLEU	BERTScore	Semb	RadGraph	1/RadCliQ-v1↑	Weights
IU-Xray	RadHiera-IU-3B	582	0.2841	0.6066	0.6265	0.3127	2.2929	tar.gz
ReXGradient/CXR	RadHiera-Rex-3B	9994	0.3059	0.5302	0.5052	0.3387	1.5373	tar.gz
MIMIC-CXR	RadHiera-MIMIC-3B	2347	0.2590	0.4919	0.4679	0.2809	1.2153	tar.gz

F1 and consistency

Dataset	Checkpoint	CheXbert F1	Consistency
IU-Xray	RadHiera-IU-3B	0.5191	0.9141
ReXGradient/CXR	RadHiera-Rex-3B	0.3760	0.8910
MIMIC-CXR	RadHiera-MIMIC-3B	0.3940	0.8509

See results/f1_consistency_summary.tsv, eval/f1_consistency/protocol_presets.json, and eval/compute_chexbert_f1_consistency_paperlike.py.

Repository contents

• weights/: released RadHiera 3B LoRA checkpoints.
• eval/: dataset-specific evaluation entry points and shared metric converters.
• train/: RadHiera MS-Swift reward plugin for GRPO training; place plugin.py in the Swift plugin workflow to enable training.
• results/: compact metric summaries for the released checkpoints.
• index.html: project page.

Evaluation

python eval/eval_mimic.py --jsonl path/to/mimic_infer.jsonl --out-dir outputs/mimic --force
python eval/eval_rex.py --jsonl path/to/rex_infer.jsonl --out-dir outputs/rex --force
python eval/eval_iu_xray.py --jsonl path/to/iu_infer.jsonl --out-dir outputs/iu --force

Dataset access and preprocessing follow the original dataset licenses and policies.