Carving Radiology Reports at Their Evidential Joints: Anatomical Region Sets as Generation Units

Jinwei Yang; Junhao Chen; Jianwei Zhao; Zhonghan Xu; Hong Cheng; Ling Wang

cs.CLOct 21, 2026

Carving Radiology Reports at Their Evidential Joints: Anatomical Region Sets as Generation Units

Authors: Jinwei Yang, Junhao Chen, Jianwei Zhao, Zhonghan Xu, Hong Cheng, Ling Wang

Organizations: University of Electronic Science and Technology of China

Abstract

Organizing radiology reports into independently generated units offers a way to generate fine-grained clinical content and bind it to visual evidence. However, existing methods typically adopt generation-unit structures not directly specified in natural reports or their annotations, introducing additional modeling compromises. We propose RadCarve, which instead defines generation units directly from the anatomical region sets in existing statement-region correspondences, unifying unit formation with visual evidence selection. RadCarve's planner uses composition-specialized query banks to jointly predict unit existence and region composition from the image, iteratively refining both with regional evidence aggregated according to current membership predictions. The generator independently produces each unit's text conditioned on the joint visual evidence of its predicted region set. Compared with strong baselines, RadCarve improves CXRFEScore by 9.4% in unit generation given region sets and achieves 2.4 times the end-to-end binding F1. It also achieves leading report-generation results on Chest ImaGenome and PadChest-GR, using fewer than one-sixth as many parameters as large-scale baselines.

Figures & tables

Figure 1: Overview of RadCarve. The planner predicts region sets $R_{t}$ , each defining a generation unit and its visual evidence. The generator independently produces unit texts $y_{t}$ , which are combined into a complete report. Snowflakes indicate frozen components.

Figure 2: Region-set planner with composition-specialized query banks and iterative composition refinement. The

\times 5

denotes five feedback updates across the six-layer decoder; intermediate layers are omitted. Final predictions specify unit existence and region composition.

Figure 3: Evidence-composition interventions on 86,011 pairs. Dots show per-token log-likelihood gains of Full set over each control for 5,000 randomly sampled pairs per condition. Densities, boxes (quartiles and medians), and diamonds (means) use all pairs. Exact scores and bootstrap confidence intervals are reported in Table 15 .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Regional input	Training text
RGRG	{right lung}	The lungs are clear. No signs of pneumonia or CHF.
	{left lung}	The lungs are clear. No signs of pneumonia or CHF.
	{cardiac silhouette}	No signs of pneumonia or CHF. Cardiomediastinal silhouette appears normal.
	{mediastinum}	Cardiomediastinal silhouette appears normal. The aorta is slightly unfolded.
	{upper mediastinum}	Cardiomediastinal silhouette appears normal.
	{aortic arch}	The aorta is slightly unfolded.

Appendix

Table 7: Comparison of training-unit constructions in a Chest ImaGenome example. RadCarve w/o pruning shows only the two units affected by pruning.

Component	Configuration
Anatomical detector	ResNet-50 Faster R-CNN; 29 regions; $512\times 512$ input; 1024-dimensional ROI features
Rad-DINO	12-layer ViT; width 768; patch size 14; $448\times 448$ input; CLS and $32\times 32$ patch tokens
Planner region encoder	29 region tokens; width 256; 2 layers; 4 attention heads
Planner query decoder	Width 256; 6 layers; 4 attention heads; 12 queries (Broad/Cross/Specific: 8/2/2); iterative refinement enabled
Language model	GPT-2-medium; 24 layers; width 1024; 16 attention heads
Generator visual interface	1024-dimensional region and region-set tokens; trainable gated visual cross-attention in all 24 language-model layers

Appendix

Table 10: Architecture configuration of the full RadCarve model.

Hyperparameter	Setting
Optimizer	AdamW
Learning rate	$5\times 10^{-5}$
Weight decay	$0.01$
Batch size	$8$
Gradient accumulation	$9$ batches
Maximum target length	$192$ tokens

Appendix

Table 12: Generator-training hyperparameters.

Method	Latency (s/report) $\downarrow$
ChEX	1.33
RGRG	1.98
MAIRA-2	4.21
MAIRA-2 (grounded)	5.12
CURE	65.42
PriorRG	1.41

Appendix

Table 13: Mean complete-report inference latency on an A100 with image batch size one.

Figure 4: Qualitative examples from the test set. Colored boxes mark detected anatomical regions, with matching colors in the selected predicted region sets. Generated unit texts are shown alongside the complete reference report.

Method	P ${}_{\mathrm{micro}\text{-}14}$ $\uparrow$	R ${}_{\mathrm{micro}\text{-}14}$ $\uparrow$	Units / report	Duplicate rate $\downarrow$	Final length / Ref.
RadCarve	0.564	0.507	5.60	0.139	0.981
w/o Set Binding	0.556	0.393	9.30	0.310	0.801

Appendix

Table 17: Clinical content recovery and generation statistics after singleton decomposition. Duplicate rate measures exact sentence repetitions before deduplication; Final length / Ref. is the final-to-reference report word-count ratio. Best results are bolded.

Broad/Cross/Specific	Set IoU $\uparrow$	[email protected] $\uparrow$	[email protected] $\uparrow$	[email protected] $\uparrow$	Pred/GT
$4/2/2$	0.455	0.615	0.524	0.566	0.853
$8/2/2$ (default)	0.567	0.551	0.642	0.593	1.165
$12/2/2$	0.579	0.535	0.655	0.589	1.225
$10/1/1$	0.573	0.549	0.648	0.595	1.180
$6/3/3$	0.506	0.590	0.577	0.583	0.979

Appendix

Table 21: Training-unit distribution and query allocation by composition type. p99 denotes the 99th percentile of units per report; Coverage is the fraction of reports whose target count does not exceed the bank’s query capacity.

Comparison	Metric	Difference	95% CI	$p$
Given anatomical region sets
RadCarve $-$ RGRG	Region IoU	$+0.0014$	$[+0.0007,+0.0021]$	$<0.001$
	Pair Pass	$+0.0053$	$[+0.0039,+0.0067]$	$<0.001$
	CXRFES	$+0.0496$	$[+0.0482,+0.0510]$	$<0.001$
	CXRFES ≥2	$+0.0337$	$[+0.0322,+0.0353]$	$<0.001$
	F1 ${}_{\mathrm{mic}\text{-}5}$	$+0.0639$	$[+0.0601,+0.0676]$	$<0.001$

Appendix

Table 25: Selected paired differences for generation given region sets (Table 1 ) and end-to-end binding recovery (Table 2 ).

Baseline ( $\mathrm{RadCarve}-$ )	Metric	Difference	95% CI	$p$
In-domain report generation
RGRG	B1	$+0.0137$	$[+0.0127,+0.0148]$	$<0.001$
RGRG	B4	$+0.0044$	$[+0.0034,+0.0053]$	$<0.001$
RGRG	METEOR	$+0.0026$	$[+0.0020,+0.0031]$	$<0.001$
R2Gen-Mamba	R-L	$-0.00003$	$[-0.0013,+0.0012]$	0.962
RGRG	CXRFES	$+0.0038$	$[+0.0031,+0.0045]$	$<0.001$

Appendix

Table 26: Paired differences against the highest-scoring baseline for each of the eight in-domain metrics (Table 3 ) and the four external metrics (Table 4 ).

Radiology Report Generation

Radiology report generation uses medical images and clinical context to produce structured, clinically accurate findings and impressions.

See all 159 papers

Harrison.Rad 1.5 Technical Report: A radiology foundation model that can draft reports from images, priors and clinical context
Jul 7, 2026Suneeta Mall, Vladimir Nekrasov, Ashnil Kumar +5Radiology Report Generation Radiology VLMs
CliQ-RRG: Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation
Oct 9, 2026Chandravardhan Singh Raghaw, Shahid Shafi Dar, Arnav Nishith Sharma +2Radiology Report Generation Medical VLMs
Indian Institute of Technology Indore, Indore, India · University of Texas at Dallas, USA · University of Wisconsin - Madison, USA
Learning to See Locally and Align Clinically with Pathology Semantics for Radiology Report Generation
Jul 31, 2026Xuan Cuong NgoRadiology Report Generation VLM Robustness

Jul 7, 2026J/K move · Enter open · S save

Harrison.Rad 1.5 Technical Report: A radiology foundation model that can draft reports from images, priors and clinical context

Suneeta Mall, Vladimir Nekrasov, Ashnil Kumar, Sajith Karunasena+4

Radiology Report Generation Radiology VLMs Multimodal Large Language Models Radiology Report Evaluation

Open paper AskPDF arXiv

Imaging demand is growing faster than the radiology workforce can expand, and reporting backlogs cannot be resolved through training and recruitment alone. The most direct opportunity is reducing the time and effort radiologists spend producing reports, a task that requires interpreting images, integrating clinical history and prior studies, and drafting structured findings. We present Harrison.Rad 1.5 (HR1.5), a radiology-specific multimodal large language model that accepts interleaved text and visual inputs and generates structured and unstructured text across plain-film radiology, spanning computed radiography, chest, musculoskeletal, abdominal, spine, and pelvic x-rays, and mammography. HR1.5 is trained through a three-stage pipeline: domain adaptation of a base language model on radiology reports, contrastive vision-encoder training with curriculum-based hard negatives on ~6 million image-report instances, and visual-question-answering fine-tuning on multi-turn conversations. We evaluate it with a Findings-Diagnosis scoring framework that extends RadGraph-XL entity extraction with ontology-based synonym matching and polarity-contradiction detection, benchmarked on RadBench, a simulated FRCR 2B Short Case examination scored against Angoff-method thresholds, ReXGradient, and internal multi-modality datasets. HR1.5 is the only system evaluated to meet the simulated FRCR passing standard and achieves the highest accuracy on closed-format clinical questions, across anatomical regions, on internal multi-body-part and mammography reporting, and on the primary clinically-aligned score for public chest reporting. We further examine explainability and model behaviour, including question-sensitive Grad-CAM heatmaps, attention analysis, and confidence estimation, to support responsible future evaluation toward clinical use, and a framework for clinically grounded assessment of report quality.

Abstract

Radiology Report Generation

Explore similar work

Harrison.Rad 1.5 Technical Report: A radiology foundation model that can draft reports from images, priors and clinical context

CliQ-RRG: Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation

Learning to See Locally and Align Clinically with Pathology Semantics for Radiology Report Generation