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

Motivation

• Radiology reports have a lower tolerance for errors because inaccurate locations or clinical descriptions can have serious consequences.The motivation is to generate fine-grained content and associate it with visual evidence.Source: §1 · p. 1

Problem

• Existing methods use unit structures not directly specified in natural reports or annotations, including predefined topics, individual regions, or merged region sets.These choices can require report rewriting, rely on externally specified units, fragment evidence for multi-region statements, or yield coarser supervision.Source: §1 · p. 1

Approach

• RadCarve jointly predicts unit existence and region composition from images, rather than relying on units supplied in advance.Composition-specialized query banks and iterative composition refinement support this joint prediction.Source: §1 · p. 2
• RadCarve defines generation units directly from anatomical region sets in statement–region correspondences, unifying unit formation with visual evidence selection.Each predicted set specifies both a unit and the regional evidence used to generate its text.Source: §1 · pp. 1–2

Results

• Experiments report more accurate recovery of multi-region bindings, leading in-domain results on Chest ImaGenome, and the highest scores on all three clinical metrics in PadChest-GR evaluation.The paper also reports that joint regional evidence supports more complete clinical content.Source: §1 · p. 2

2.1 Spatial Grounding and Local Evidence in Complete Report Generation

• Spatial grounding methods add localization or regional evidence, yet complete reports are still generated through a single continuous autoregressive process.Controlled experiments report that the gap between real- and blank-image performance narrows as more preceding report text is supplied, so spatial information alone does not ensure each statement is determined by its visual evidence.Source: §2.1 · p. 2

2.2 Generation Units in Radiology Report Generation

• Topic-based methods use latent intentions or predefined schemes, while anatomy-based methods select single regions or split cross-region statements into single-region units.Natural sentence boundaries need not match topic boundaries, and manually annotated PadChest-GR includes single-finding sentences associated with multiple regions.Source: §2.2 · p. 2
• Dalla Serra et al. (2023) merge overlapping region sets and associated sentences, which can expand transitively and produce coarser supervision; inference uses externally supplied subsets.CURE uses predefined anatomical queries, while ChEX filters queried regions using image-based classification.Source: §2.2 · pp. 2–3
• RadCarve predicts unit existence and region composition from the image, defining units directly from statement–region sets without single-region projection or expansion through shared regions.Figure 1 shows predicted sets defining both units and visual evidence, followed by independent unit-text generation and report assembly.Source: §2.2 · p. 3, Figure 1 · p. 3

3.1 Overview

• RadCarve uses a planner to predict nonempty region sets, each defining a generation unit and its visual evidence, then independently generates unit text from set-level and member-region features.The unit texts are combined into a complete report.Source: §3.1 · p. 3

3.2 Generation Units and Supervision Construction

• Each distinct nonempty statement–region set defines a source unit, and its associated statements are concatenated as the text target after redundant parent regions are pruned.Pruning retains finer-grained regions when both a specific region and its broader parent occur in a set; the resulting set and text target supervise the planner and generator, respectively.Source: §3.2 · p. 3

3.3 Region-Set Planning

• The planner updates each candidate unit’s set-level representation to refine unit existence and region composition from visual memory.Membership probabilities weight regional evidence, and the aggregate updates the representation through a gated residual.Source: §3.3 · pp. 4–5
• Composition-specialized query banks divide sets into Broad, Specific, and Cross types, while iterative refinement aggregates regional evidence according to current membership predictions.The banks share the visual encoder and Transformer decoder but use separate prediction heads; unmatched queries are supervised to predict nonexistence.Source: §3.3 · pp. 4–5, Figure 2 · p. 4

3.4 Unit-Conditioned Text Generation

• For each predicted set, the generator conditions on a set-level token and tokens for its member regions, then independently decodes unit text.A frozen GPT-2-medium backbone with trainable gated cross-attention is trained on ground-truth and predicted sets; inference assembles decoded unit texts after sentence-level deduplication.Source: §3.4 · p. 5

4.2 Generation Units and Evidence Binding

• RadCarve outperforms all baselines, including CURE, on CXRFES and all three clinical F1 metrics when generating text given anatomical region sets.Its CXRFES lead persists on multi-region pairs despite similar localization performance to RGRG.Source: §4.2 · p. 6, Table 1 · p. 6
• 0.451 binding F1@0.5 for RadCarve, compared with 0.185 for RGRG and 0.089 for ChEX, in end-to-end binding recovery.RadCarve also achieves higher Multi IoU than RGRG despite similar Single IoU.Source: §4.2 · p. 6, Table 2 · p. 6
• Full set yields positive per-token log-likelihood gains over Best singleton and Random same-size, while removing members weakens support for target text.The advantage holds for nearly all evaluated pairs across all three controls, indicating use of joint evidence from the selected region set.Source: Figure 3 · p. 6, §4.2 · p. 7

4.3 Report Generation Performance

• RadCarve leads Chest ImaGenome report generation on CXRFES, all three clinical F1 scores, BLEU-1, BLEU-4, and METEOR, and ties for highest ROUGE-L.It uses 0.761B inference parameters, versus 6.880B for MAIRA-2 and 5.010B for CURE.Source: §4.3 · p. 7, Table 3 · p. 7
• On PadChest-GR without target-domain training or fine-tuning, RadCarve achieves the highest scores on all three clinical metrics.MAIRA-2 and CURE are excluded because they used PadChest report supervision during training.Source: §4.3 · p. 7, Table 4 · p. 7
• On all-black input, RadCarve generates one unit versus 5.6 on real images, and its text receives no CheXbert observation labels.This contraction, together with real-image unit-recovery results, suggests the planner adjusts generation in response to image content.Source: §4.3 · p. 7

4.4 Ablation Studies

• Without set binding, Rmicro-14 falls from 0.507 to 0.393 while Pmicro-14 changes from 0.564 to 0.556, and exact duplicate sentence rate rises from 0.139 to 0.310.Disabling pruning reduces all metrics; replacing only the unpruned model’s generator with the pruned-supervision generator raises F1micro-14 from 0.482 to 0.524.Source: §4.4 · p. 8, Table 5 · p. 8
• Without composition refinement, region-only Set IoU falls from 0.567 to 0.553 despite similar predicted unit counts; homogeneous query banks also reduce report metrics.The region-only F1@0.5 drops by 0.064 for Cross and 0.062 for Specific sets, versus 0.008 for Broad sets, without specialized banks.Source: §4.4 · p. 8, Table 6 · p. 8

• RadCarve derives generation units directly from anatomical region sets in existing statement–region correspondences, making each set both a unit definition and evidence selection.Source: §5 · p. 8
• Evaluations show improved recovery of multi-region bindings and support joint use of related regions for more complete clinical content.Source: §5 · p. 8
• These benefits extend to complete-report generation, with strong in-domain and external performance.Source: §5 · p. 8
• The conclusion is that visual evidence composition should inform generation-unit definition, not only condition text generation.Source: §5 · p. 8

• Automatically constructed Chest ImaGenome statement–region correspondences may include content not strictly related to their associated regions, and pruning does not resolve all region–text mismatches.Source: Limitations · p. 9
• Experiments use one current frontal radiograph without prior examinations, lateral views, or clinical context, limiting assessment of interval changes and verification of comparative statements.Source: Limitations · p. 9

Ethical Considerations

• RadCarve is intended for assisted-report-generation research; clinical use requires further validation and radiologist oversight.Controlled-access MIMIC-CXR and Chest ImaGenome data were used under their data use agreements, and PadChest-GR was used for external evaluation under its terms of use.Source: Ethical Considerations · p. 9

A.1 Statement–Region Correspondences and Supervision Construction

• Each distinct nonempty statement-associated region set defines a training unit, while its associated statements form the text target.Statements are built by merging phrases with the same normalized text, with their associated regions collected into a set.Source: §A.1 · p. 12
• RadCarve keeps four separate region-set units in the example, accommodating cross-region descriptions and localized findings.By comparison, RGRG projects multi-region correspondences onto individual regions, while Dalla Serra et al. merge overlapping sets through shared regions.Source: §A.1 · p. 12

A.2 Anatomical Redundancy Pruning

• The 21 redundancy rules remove a parent only when its corresponding child is also in the same set, preserving parent-only and bilateral lung sets.For the dataset-specific overlap rules, hilar boxes are almost entirely covered by ipsilateral lung boxes; hemidiaphragm boxes cover 86.32%–92.14% of abdomen boxes, and removing abdomen changes targets in 6.75%–9.73% of co-occurring units.Source: §A.2 · p. 12

A.3 Unit and Supervision Statistics

• 58.12% of training region sets but 45.78% of text targets change after pruning, while mean set cardinality falls from 3.015 to 2.125.Multi-region units remain the majority: 71.31% of training units contain at least two regions after pruning.Source: §A.3 · p. 12, Table 9 · p. 14

B.1 Model Configuration and Feature Representations

• The planner combines detector region features with gated local Rad-DINO features, using image-conditioned queries and patch tokens in its visual memory.Each query maintains a candidate unit’s set-level representation, updated across decoder layers with visual memory and composition feedback.Source: §B.1 · p. 15

B.2 Region-Set Planner Details

• Composition-specialized Broad, Cross, and Specific query banks jointly predict unit existence and membership, with iterative refinement sharing parameters across feedback updates.The 12-region subset used to distinguish composition types targets regions with low training frequency and low regional F1 despite detector recall of at least 80%.Source: §B.2 · pp. 15–16

B.3 Generator Training and Inference

• Predicted-set batches increase from 1/9 in the first epoch to 4/9 in the selected fourth-epoch checkpoint, while validation uses predicted sets throughout.Predicted sets are matched to ground-truth units within composition types, and overlapping matched pairs contribute to text loss.Source: §B.3 · p. 17
• At inference, each predicted unit is independently decoded with four-beam search, and nonempty unit texts are concatenated before duplicate-sentence removal.The second removal stage retains the longer sentence when a pair has BERTScore F1 greater than 0.9.Source: §B.3 · p. 17

C.1 Dataset Statistics

• Complete-report evaluation uses all 32,711 Chest ImaGenome test images, while external PadChest-GR evaluation uses 915 official test samples without target-domain training or fine-tuning.PadChest-GR provides finding-level text–box annotations, so it is used for complete-report evaluation.Source: §C.1 · p. 17

C.3 Unit Generation Given Anatomical Region Sets

• The given-set evaluation covers 156,540 text–region-set pairs from 32,568 images, with methods given regions but neither target text nor ground-truth boxes.Text receives credit only when every target region is localized at IoU ≥0.5; failed pairs receive empty candidate text.Source: §C.3 · p. 18
• On all-black inputs, RGRG and RadCarve’s complete-pair rates fall from 24.74% and 22.13% to 0.55% and 2.77%, respectively.CURE still passes on 29.11% of pairs, while ChEX’s 17.98% rate is close to its 19.74% real-image rate; passing does not establish visual support.Source: §C.3 · pp. 18–19

C.4 End-to-End Binding Recovery

• End-to-end binding recovery is evaluated on 31,271 images without supplied target region sets, matching unit texts before evaluating their region sets by Jaccard IoU.The region-only protocol instead matches predicted and target sets by total Jaccard IoU.Source: §C.4 · p. 19

C.5 Qualitative Examples

• Figure 4 illustrates single- and multi-region units covering relevant findings, including effusions, atelectasis, cardiomegaly, and normal contours.One cardiac unit describes cardiomegaly as moderate, whereas the reference says mild enlargement.Source: §C.5 · pp. 19, 21, Figure 4 · p. 20

C.6 Evidence Composition Intervention

• The evidence-composition intervention holds the image, generator, target text, and regional features fixed while changing only the input region set.It compares the full set with best-singleton, leave-one-out, and random same-size controls, scoring per-token negative log-likelihood.Source: Table 15 · p. 20, §C.6 · p. 21

C.7 Report-Level Metrics

• Complete-report comparisons use common test samples, reference reports, and scoring procedures, with CXRFES and CheXbert metrics defined alongside text-overlap scores.Baseline training splits overlap the processed Chest ImaGenome test set and may favor some baselines; only 556 images (1.7%) also belong to the official MIMIC-CXR test split.Source: §C.2 · p. 18, §C.7 · p. 21

C.8 All-Black Input

• On all-black inputs, RadCarve contracts from 5.6 average real-image units to one unit, whose remaining sentence receives no CheXbert observation labels.The planner predicts a single left-lung unit despite six detector outputs; the sentence makes an unsupported prior-study comparison.Source: §C.8 · p. 22

D.1 Generation-Unit Construction Ablations

• Removing set binding preserves anatomical coverage but decomposes grouped regions into independently generated single-region units.The resulting diagnostics show a substantial clinical-recall drop, a smaller precision decrease, more repeated sentences, and shorter reports.Source: §D.1 · p. 22, Table 17 · p. 23
• The unpruned model disables pruning during training and inference, while Hybrid replaces only its generator with one trained under pruned supervision.The comparison retains the unpruned planner outputs, region sets, and region-set tokens.Source: §D.1 · p. 22
• 44.17% of predicted sets retain parent–child redundancy without pruning, while Hybrid recovers much of the clinical-score gap on those same unpruned sets.This indicates that supervision constructed after pruning contributes substantially to the improvement.Source: §D.1 · p. 22
• The ablation tables also report region-set recovery under region-only matching, Set IoU across decoder layers, and recovery by target composition type.The supplied captions identify these evaluation scopes but do not provide the numerical comparisons.Source: Table 18 · p. 23, Table 19 · p. 23, Table 20 · p. 23

D.2 Planner Ablations and Recovery Diagnostics

• F1@0.5 is 0.593 for RadCarve, 0.336 for RGRG, and 0.168 for ChEX in region-set recovery; refinement and specialized banks also improve recovery.RadCarve has higher Set IoU than the no-refinement variant at every decoder layer. Removing specialized banks lowers F1@0.5 from 0.593 to 0.577.Source: §D.2 · pp. 23–24, Table 21 · p. 24, Table 22 · p. 24

D.3 Bank Capacity and Allocation Sensitivity

• At fixed total capacity 12, 10/1/1 yields F1 0.595 versus 0.593 for 8/2/2, while 6/3/3 raises precision but lowers Set IoU from 0.567 to 0.506.With Cross and Specific held at two queries, 4/2/2 lowers Pred/GT to 0.853 and recall to 0.524; 12/2/2 raises them to 1.225 and 0.655, but has lower F1 than 8/2/2.Source: §D.3 · p. 25
• Under 10/1/1, Cross produces almost no valid units despite slightly higher overall F1; expanding Cross and Specific to three queries reduces both banks’ output and F1.The authors characterize 8/2/2 as a balanced allocation, not the best on every metric.Source: §D.3 · pp. 24–25, Table 23 · p. 24

D.4 Generator Component Ablations

• Removing the region-set token or local Rad-DINO fusion modestly changes text-overlap and CXRFES scores, but lowers clinical F1 in both variants.The planner is unchanged in both variants; without local Rad-DINO fusion, the generator uses detector-derived region features without its local Rad-DINO branch.Source: Table 24 · p. 24, §D.4 · p. 25

E.1 Paired Bootstrap Protocol

• The paired bootstrap resamples images/reports, keeps all text–region-set pairs from each image together, and uses the same sampled indices for compared methods.It uses 10,000 resamples; 95% confidence intervals for paired differences are the 2.5th and 97.5th bootstrap percentiles.Source: §E.1 · p. 25

E.2 Generation Units and Evidence Binding

• Against RGRG, the Single IoU difference is small and its percentile confidence interval crosses zero, so the authors do not claim a clear single-region recovery advantage.The centered-bootstrap p-value is 0.0475; the reported end-to-end results show gains in overall binding recovery and multi-region IoU.Source: §E.2 · p. 25, Table 25 · p. 26

E.3 Complete-Report Generation

• On PadChest-GR, gains in both micro-F1 metrics are significant, while CXRFES and F1ex-14 differences against their named baselines are not statistically significant.The comparisons use the highest-scoring baseline for each metric; the supplied passages do not state the numerical differences.Source: §E.3 · p. 25, Table 26 · p. 26

E.4 Key Ablation Comparisons

• Specialized-bank tests compare the full model with w/o Specialized Banks within each composition type, not differences in improvement across types.Source: §E.4 · p. 27
• All listed clinical F1 differences for the main and generator-component ablations have positive 95% CIs (p < 0.001).Source: §E.4 · p. 27
• Hybrid is compared with w/o Pruning to quantify generator-change recovery on unpruned inputs and with full RadCarve to measure the remaining gap.Source: §E.4 · p. 27
• Table 28 covers singleton diagnostics, overall planner recovery, and composition-type recovery; duplicate rate is lower-is-better and other metrics are higher-is-better.Singleton clinical-recall and duplicate-rate rows characterize decomposition; region-only matching evaluates planner recovery without generated text, and composition comparisons report positive IoU and F1@0.5 differences.Source: §E.4 · p. 27, Table 28 · p. 28