Chest X-ray reports now generate by anatomical region sets: RadCarve leads multi-region binding with under one-sixth the parameters
Synopsis
Chest X-ray report generation now defines its generation units directly as the anatomical region sets that statements are linked to in report annotations, rather than preset topics or single regions. RadCarve has a planner predict from the image which units exist and which regions each contains, then generates each unit's text independently from the joint visual evidence of that region set. On Chest ImaGenome, given region sets it improves CXRFEScore by 9.4% over strong baselines, and its end-to-end binding F1@0.5 is 0.451 versus 0.185 for RGRG and 0.089 for ChEX, with 0.761B inference parameters.
Interpretation
Chest X-ray reports are no longer split into preset topics or individual anatomical regions; generation units follow the anatomical region sets that statements are linked to in report annotations, so one unit can span several regions. Earlier unit-based methods either rewrote reports to align with preset topics, projected multi-region statements onto single regions, or merged overlapping region sets through shared regions into coarser supervision. In the Chest ImaGenome training data, 85.2% of region sets contain two or more regions, so multi-region units are the common case rather than an exception.
A planner works from the image to decide jointly whether a candidate unit exists and which anatomical regions compose it, with the two predictions refining each other during decoding. Queries are organized into Broad, Cross and Specific composition-specialized banks with separate prediction heads, and each decoder layer feeds back regional evidence aggregated by the current membership predictions into the unit representation. End-to-end evaluation on 31,271 test images without supplying target region sets gives RadCarve a binding F1@0.5 of 0.451, against 0.185 for RGRG and 0.089 for ChEX.
The generator conditions on the set-level representation of a predicted region set plus its member-region features, produces each unit's text independently, and assembles the units into a complete report. On 86,011 correctly localized multi-region pairs, replacing the input region set with the best singleton, deleting members one at a time, or using equally sized sets of nontarget regions all lower the per-token log-likelihood of the target text. Complete-report generation reaches the highest CXRFEScore and all three clinical F1 scores on Chest ImaGenome, and the highest scores on all three clinical metrics on the 915-sample PadChest-GR test set without target-domain training.
Perspective
Evaluation covers complete-report generation on the 32,711-image processed Chest ImaGenome test set, 156,540 given-region-set text pairs, and end-to-end binding on 31,271 images; external testing uses the official 915-sample PadChest-GR test set without target-domain training or fine-tuning. All experiments use only the current single frontal chest radiograph, without prior examinations, lateral views or clinical context. The authors position the system as research on assisted report generation; clinical use would require further validation and radiologist oversight.
Supervision comes from automatically constructed statement-region correspondences in Chest ImaGenome, which may include content not strictly related to the associated regions; anatomical redundancy pruning handles specific parent-child redundancies but does not resolve all region-text mismatches. A single current frontal image provides no longitudinal evidence for assessing interval changes, and comparative statements the model generates cannot be verified from the supplied input alone. In external evaluation, the CXRFEScore difference against PriorRG and the F1 difference against R2Gen-Mamba are not statistically significant.