Structured Radiology Report Generation (S-RRG)

• Structured Radiology Report Generation is a process of creating radiology reports with explicit anatomical and clinical structures, ensuring machine-readability.
• It employs diverse methods such as prompt-enforced sectioning, ontology alignment, and procedural pipelines to improve accuracy and mitigate hallucinations.
• Evaluation of S-RRG focuses on clinical fidelity, anatomical correctness, and structured data reusability, driving advances in multimodal report generation.

Searching arXiv for papers on structured radiology report generation and the cited methods to ground the article in current literature. Structured Radiology Report Generation (S-RRG) is the radiology report generation setting in which outputs are not emitted as a single unconstrained narrative block, but are organized into explicit, machine-parseable structures such as section boundaries, anatomy-wise findings, key–value pairs, and optionally ontology-constrained representations. In the cited literature, S-RRG is motivated by the need for clinically useful reports whose content is anatomically grounded, structurally consistent, and suitable for downstream coding, registry construction, longitudinal tracking, and other forms of clinical data reuse (Gao et al., 14 Nov 2025). Recent work formulates this structure at several levels: fixed report sections and organ headers, structured finding assertions, JSON schemas, lesion- and evidence-linked objects, or hierarchical entity–attribute representations (Delbrouck et al., 30 May 2025).

1. Conceptual scope and representational forms

S-RRG is not a single output format but a family of representational commitments. One line of work defines structured outputs through standardized sections and constrained discourse. In chest X-ray SRRG, reports are generated with section headers such as Exam Type, History, Technique, Comparison, Findings, and Impression; Findings are further organized under a closed set of anatomical headers, and Impression is formatted as a numbered list ordered by clinical importance (Delbrouck et al., 30 May 2025). A closely related SRRG schema structures Findings by anatomical regions such as Lungs and Airways, Pleura, Cardiovascular, Hila and Mediastinum, Tubes, Catheters, and Support Devices, Musculoskeletal and Chest Wall, and Other, while Impression is represented as numbered diagnostic syntheses (Kang et al., 1 Oct 2025).

A second line of work treats structure as an ontology-aligned decomposition of report content into entities and attributes. ReportQA formalizes a structured report as a set of clinical entities and their attributes, including presence, location, size, severity, chronicity, certainty, and other dimensions within radiologist-guided knowledge trees (Shi et al., 13 Jun 2026). S-RRG-Bench similarly encodes each positive finding as a record with disease name, severity level, probability, and anatomical location, and each negative finding as an explicit disease-name entry in JSON (Li et al., 4 Aug 2025). This makes structure an intrinsic property of the generated artifact rather than a post hoc formatting convention.

A third formulation places structure in the internal reasoning or evidence model. CogRad defines “structure” procedurally as a four-stage pipeline—global triage, focused investigation, structured drafting, and verification—while also maintaining disease/region-level structure through learned regions and explicit 14-disease probabilities (Khan et al., 4 Jul 2026). Evidence-linked reporting extends this idea further by treating the report as a human-supervised layer that links report statements to measurements, segmentations, image references, prior comparisons, lesion identity, uncertainty qualifiers, and standardized terminology (Kazemzadeh et al., 24 May 2026).

A recurrent misconception is that S-RRG is equivalent to adding section headers to otherwise unchanged free text. The cited work suggests a broader view. Structure may reside in report sections, organ-wise grouping, JSON fields, lesion tracking identifiers, disease hierarchies, or reviewable evidence objects. This suggests that formatting alone is insufficient when anatomical grounding, temporal comparison, or downstream computability are primary requirements.

2. Core objectives: anatomical grounding, consistency, and context

Two properties recur across the literature: anatomical grounding and structural consistency. S2D-Align states these explicitly: each described finding should be tied to the correct body region, and the same finding should be consistently stated across sections and according to reporting conventions (Gao et al., 14 Nov 2025). The motivation is practical rather than stylistic. False or poorly localized findings degrade coding, registry building, longitudinal tracking, and structured extraction.

Several works show that structure without grounding remains fragile. Standard supervised fine-tuning on study-level image–text pairs can inherit templated language and weak anatomy grounding, which in turn leads to hallucinations or “no abnormality” defaults even when localized abnormalities are present (Gao et al., 14 Nov 2025). ORID addresses this by explicitly constructing organ-regional visual features and organ-aware textual descriptors, then reweighting them with an organ-importance GNN so that clinically salient organs dominate generation rather than unrelated regions (Gu et al., 2024). RIHA addresses a related problem by aligning image features and report content across paragraph, sentence, and word levels, rather than treating the report as a flat sequence (Chen et al., 30 Apr 2026).

Clinical context is another central objective. C-SRRG argues that omission of clinical indication, technique, multi-view images, and prior studies leads to systematic errors, especially temporal hallucinations such as “unchanged” or “new since prior” when no prior study is available (Kang et al., 1 Oct 2025). The proposed contextualized SRRG setup therefore inserts current metadata and prior-study blocks directly into both training and inference prompts. This is not merely an input expansion; it reframes structured reporting as a context-conditioned task in which temporal language must be licensed by explicit evidence.

The broader literature also distinguishes between structure that is reviewable and structure that is merely generative. The evidence-linked architecture is explicitly not an autonomous report finalizer. Instead, it treats reviewed reporting, provenance visualization, lesion identity management, and standards-based export as first-class constraints (Kazemzadeh et al., 24 May 2026). A plausible implication is that, in clinically deployed S-RRG, the relevant notion of correctness includes not only semantic agreement with a reference report but also traceability from each assertion to image evidence, measurements, priors, or approved structured elements.

3. Principal modeling paradigms

Recent S-RRG systems differ primarily in what kind of structure they encode and where they impose it: during conditioning, in latent alignment, at decoding time, or through post-generation validation.

Paradigm	Representative system	Structural mechanism
Progressive grounding during SFT	S2D-Align	Three-stage curriculum with reference reports, key phrases, and shared-memory adapter
Prompt-enforced sectioned SRRG	C-SRRG	Standardized prompts with section headers, organ-wise Findings, numbered Impression
Organ/hierarchy-aware alignment	ORID, RIHA	Organ-level fusion and GNN saliency; paragraph/sentence/word alignment with OT
Procedural multi-agent generation	CogRad	Scout–Investigator–Writer–Verifier pipeline with sentence-level re-examination
Schema- or label-constrained generation	LaB-RAG, dynamic-template-constrained LLM	Label-conditioned retrieval, JSON templates, controlled vocabularies
Deterministic feature-to-report rendering	BTReport	Deterministic feature extraction followed by LLM-based syntactic rendering

S2D-Align is a representative example of training-time structural induction without explicit structured decoding. It organizes supervised fine-tuning as a three-stage curriculum: coarse radiograph–report pairing, instance-level guidance through same-patient reference reports, and fine-grained guidance through RadGraph-derived key phrases. A Shallow-to-Deep Memory Adapter with shared learnable memory queries integrates image, reference-report, and key-phrase features into a common conditioning space. During inference, only image conditioning is used, so the model retains standard runtime while benefiting from shallow-to-deep grounding learned during training (Gao et al., 14 Nov 2025). Although it outputs free-form Findings text rather than explicit sections, the paper positions it as a useful substrate for downstream structured extraction.

C-SRRG enforces structure more directly at the prompt level. It fine-tunes medical multimodal LLMs such as CheXagent-3B, MedGemma-4B, and Lingshu-7B with prompts that require structured Findings or Impression outputs, and it concatenates current and prior multimodal context into a single multimodal token sequence (Kang et al., 1 Oct 2025). No external constraint module is introduced; the structure is enforced through instruction prompts, fixed section headers, and benchmark design. This approach targets clinically aligned sectioning while also reducing temporal hallucinations through explicit prior conditioning.

ORID and RIHA encode structure inside the alignment machinery. ORID produces organ-aware visual features via segmentation masks, organ-wise textual descriptors from an instruction-tuned LLaVA-Med-RRG, organ-based cross-modal fusion, and graph-based organ-importance coefficients that suppress cross-organ noise (Gu et al., 2024). RIHA instead uses a Visual Feature Pyramid and Text Feature Pyramid, with optimal transport losses aligning visual and textual units at paragraph, sentence, and word levels, and Relative Positional Encoding in the decoder to improve token-level coherence (Chen et al., 30 Apr 2026). These systems do not necessarily impose explicit section headers, but they treat report structure as a hierarchical or organ-specific latent variable rather than a flat sequence-generation problem.

CogRad moves structure into procedural control. Its Scout agent discovers regions with slot attention and predicts disease-level triage scores; the Investigator allocates attention to suspicious regions; the Writer constructs a disease-gated visual prefix for the LLM; and the Verifier applies a training-time visual entailment loss and an inference-time sentence-level re-examination loop that can regenerate the report if a sentence is judged insufficiently grounded (Khan et al., 4 Jul 2026). This introduces a form of structured self-correction not present in single-pass generators.

A different family of methods constrains generation through explicit schemas, label sets, or templates. LaB-RAG converts a chest X-ray into discrete CheXpert-derived labels via logistic regressions over frozen BioViL-T embeddings, uses those labels to boost retrieval, and then prompts an instruction-tuned LLM to generate section-specific reports; the same labels can also be used as structured primitives for JSON output (Song et al., 2024). The dynamic-template-constrained LLM for lung cancer screening report conversion goes further by integrating a dynamic JSON template and controlled vocabularies directly into decoding, constraining every slot to predefined value sets and reporting neither formatting errors nor content hallucinations (Niu et al., 2024).

BTReport represents the strongest form of explicit measurement grounding. It deterministically extracts clinically relevant brain tumor features from mpMRI—including lesion sizes, volumes, midline shift, ventricular status, and modified VASARI descriptors—and uses the LLM only for syntactic structuring and narrative formatting of a Findings section (Rivera et al., 17 Feb 2026). This separation of deterministic interpretation from language realization is structurally different from end-to-end image-to-text generation and is presented as a mechanism for interpretability and hallucination reduction.

4. Datasets, supervision sources, and benchmark construction

Structured reporting research has depended heavily on re-curating free-text corpora into structured targets or auxiliary supervision. Several resources are particularly prominent.

Resource	Scope	Structured elements
C-SRRG	Chest X-ray Findings and Impression tasks	Anatomical-region Findings and numbered Impression with clinical context
SRRG / StructUtterances	Chest X-ray	Sectioned reports, utterance-level disease labels, 55-label hierarchy
MIMIC-STRUC	Chest X-ray	JSON with disease, severity, probability, location
ReportQA	CXR, brain CT, chest CT, abdominal CT	Knowledge-tree entities and attributes converted into QA pairs
BTReport-BraTS	Brain tumor MRI	Deterministic feature JSON and synthetic Findings

C-SRRG augments MIMIC-CXR and CheXpert Plus with multi-view identifiers, indication, technique, parsed comparison fields, and longitudinal patient histories. It defines two SRRG tasks: C-SRRG-Findings with 181,874/976/1,459/233 train/valid/test/test-reviewed instances and C-SRRG-Impression with 405,972/1,505/2,219/231 instances (Kang et al., 1 Oct 2025). The design is notable because context is not only present in the dataset; it is standardized into current and previous study blocks that can be consumed during both training and inference.

The SRRG dataset introduced in “Automated Structured Radiology Report Generation” restructures MIMIC-CXR and CheXpert Plus reports with GPT-4 Turbo under a strict schema. It yields SRRG-Impression with 409,927 total instances, SRRG-Findings with 184,542 total instances, and StructUtterances with 1,506,158 utterances labeled across a 55-label disease taxonomy (Delbrouck et al., 30 May 2025). The associated SRR-BERT evaluator is trained on utterance-level labels and supports scoring at leaf or upper hierarchy levels.

MIMIC-STRUC is built from MIMIC-CXR-JPG and normalizes disease mentions against a 30-disease ontology, with automated extraction of severity and probability plus GPT-4 standardization of location phrases. The resulting target format is JSON, in which positive findings carry name, probability, level, and location, and negative findings list only disease names (Li et al., 4 Aug 2025). This benchmark is designed so that both training targets and evaluation metrics operate directly on structured fields rather than on extracted text labels.

ReportQA generalizes beyond chest X-ray by covering MIMIC-CXR, CTRG-Brain, CT-RATE, and AMOS-MM. It constructs radiologist-guided entity and attribute trees, converts reports into ontology-aligned tuples, then generates and filters multiple-choice QA pairs. The final corpus contains about 660k high-quality QAs for 6,857 reports, with roughly 100 QAs per report (Shi et al., 13 Jun 2026). This is not a generation dataset in the narrow sense; rather, it redefines the evaluation target for S-RRG as structured information transfer across modalities and anatomic regions.

BTReport-BraTS augments 1,470 BraTS’23 adult glioma cases with deterministic feature JSON, midline segmentations, structured summaries, and synthetic Findings sections (Rivera et al., 17 Feb 2026). Its importance lies in showing that, outside chest radiography, S-RRG may be more naturally grounded in quantitative pipelines than in paired image–report corpora.

A separate but related resource is the lung cancer screening report corpus used for free-text-to-structure conversion. It contains 5,442 de-identified LDCT screening reports from two institutions, with 500 labeled free-text/fully-structured report pairs and a 28-feature schema spanning nodule-level, report-level, and auxiliary fields (Niu et al., 2024). This work addresses report structuring from text rather than report generation from images, but it illustrates a complementary route by which structured radiology artifacts can be produced.

5. Evaluation: from lexical overlap to structured clinical fidelity

The literature consistently argues that conventional NLG metrics are insufficient for S-RRG. BLEU, ROUGE, METEOR, and related metrics remain widely reported, but several papers explicitly note that they fail to capture clinically important entities, attributes, or structured consistency (Shi et al., 13 Jun 2026). As a result, recent evaluation frameworks emphasize structured clinical correctness, anatomy-aware section compliance, and fine-grained attribute fidelity.

One widely used family of metrics operates on extracted clinical labels or entities. S2D-Align reports BLEU-1..4 and ROUGE-L together with CheXbert-style clinical efficacy precision, recall, and F1, reaching on MIMIC-CXR $B@1=0.422$, $B@4=0.149$, $\text{ROUGE-L}=0.332$, CE Precision $=0.613$, Recall $=0.606$, and F1 $=0.608$ (Gao et al., 14 Nov 2025). CogRad evaluates BLEU, CIDEr, RadGraph F1, CheXbert F1, and entity-level hallucination and miss rates, explicitly showing that NLG gains do not necessarily translate proportionally into entity-level factuality (Khan et al., 4 Jul 2026).

Section-aware SRRG benchmarks add structure-specific metrics. C-SRRG reports F1-SRRG-BERT, built on CXR-BERT, and Category Score for organ-section headers, alongside RadGraph and standard text metrics; the reported gains under full clinical context are especially large for larger MLLMs (Kang et al., 1 Oct 2025). The SRRG benchmark based on SRR-BERT formalizes utterance-level disease scoring for structured Findings and Impression sections, evaluating both unaligned and order-sensitive aligned settings. It reports, for example, that Category F1 for Findings is consistently high while disease F1-SRR-BERT varies substantially by organ system and model (Delbrouck et al., 30 May 2025).

Two recent proposals attempt to score the full structured semantics of a report rather than just its disease mentions. S-RRG-Bench introduces S-Score, which averages a disease prediction term, P-Score, over positive and negative disease sets with a detailed-description term, D-Score, that evaluates probability, severity, and location. In the reported correlation analysis against GPT-4 ratings, S-Score achieves Kendall Tau / Spearman of $0.591 / 0.769$, exceeding BLEU-4, ROUGE, METEOR, and GREEN (Li et al., 4 Aug 2025). ReportQA introduces QAScore, defined as the harmonic mean of a positive score over QA accuracy and a negative score that exponentially penalizes false-positive answers on absent entities; on RadEvalX it shows the strongest correlation with radiologist error judgments among the compared metrics, with Pearson $|r|=0.4507$, Spearman $|\rho|=0.4612$, and Kendall $|\tau|=0.3910$ (Shi et al., 13 Jun 2026).

Template-constrained structuring work contributes a different evaluation perspective. In the lung screening conversion setting, the dynamic-template-constrained system reports cross-institutional F1 of about $B@4=0.149$0, with zero formatting errors and zero content hallucinations, and outperforms GPT-4o by $B@4=0.149$1 in the reported comparison (Niu et al., 2024). This is not an image-to-report metric, but it demonstrates that schema validity and hallucination rate can themselves be primary evaluation targets in structured reporting.

A plausible implication is that S-RRG evaluation is converging toward a layered regime. Lexical metrics remain useful for surface realization; entity and label metrics capture disease content; section metrics measure structural compliance; QA-based or attribute-aware scores assess whether the report transfers clinically actionable information. No single metric in the cited literature is presented as universally sufficient.

6. Clinical integration, misconceptions, and future directions

The cited literature repeatedly frames S-RRG as an assistive rather than autonomous technology. The evidence-linked reference architecture explicitly rejects autonomous finalization and instead positions structured reporting as a human-supervised intelligence layer integrating templates, speech-to-structure, measurement capture, segmentation, controlled AI drafting, interoperability, governance, and auditability (Kazemzadeh et al., 24 May 2026). BTReport follows a similar philosophy by separating deterministic feature extraction from language formatting, so that each narrative clause can be traced to structured measurements or categorical features (Rivera et al., 17 Feb 2026).

This emphasis addresses several common misconceptions. One is that richer structure automatically guarantees factuality. In practice, section-aware decoders and constrained formats can still degrade fluency or miss subtle findings if the underlying alignment is weak; S2D-Align, ORID, RIHA, and CogRad all treat visual grounding as a separate technical problem from structural formatting (Gao et al., 14 Nov 2025). Another misconception is that retrieval or prior context is always beneficial. C-SRRG shows that context must be coherently present or absent in both training and inference to mitigate temporal hallucinations, and retrieval-based methods can propagate mismatched anatomy or irrelevant cases when explicit grounding is insufficient (Kang et al., 1 Oct 2025).

The major limitations are also consistent across papers. S2D-Align depends on RadGraph coverage and same-patient longitudinal data for its auxiliary stages (Gao et al., 14 Nov 2025). C-SRRG is capped by model context windows and multi-image handling, and one model failed to follow the Impression format under full context (Kang et al., 1 Oct 2025). CogRad remains limited to CheXpert-14 disease supervision and still shows relatively modest RadGraph F1 on CheXpert Plus (Khan et al., 4 Jul 2026). ReportQA emphasizes that current models remain weak on fine-grained attributes such as margin, chronicity, enhancement, distribution, internal features, and secondary effects, and that report-based inference exhibits strong negative priors (Shi et al., 13 Jun 2026). S-RRG-Bench notes that automated extraction and LLM-assisted location normalization introduce their own noise and that multi-view and non-CXR extensions are not yet addressed (Li et al., 4 Aug 2025).

Several future directions recur. Ontology-constrained decoding and anatomy-slot generation are proposed as natural extensions when phrase-to-anatomy or label-to-slot mappings are available (Gao et al., 14 Nov 2025). Question-driven decoding, multi-task QA supervision, and reinforcement learning from QA-derived rewards are suggested as mechanisms for improving fine-grained entity–attribute recognition beyond report-level generation (Shi et al., 13 Jun 2026). Larger or longer-context multimodal models, improved multi-image attention, learned retrieval over PACS/EHR, and broader modality coverage recur in contextual SRRG work (Kang et al., 1 Oct 2025). Dynamic template constraints, local deployment, and standards-based export through DICOM SR, DICOM SEG, and FHIR address deployment, privacy, and interoperability concerns that purely generative benchmarks do not resolve (Niu et al., 2024).

Taken together, these studies describe S-RRG as a convergence area rather than a single architecture class. Some systems prioritize anatomically grounded generation from images, some enforce sectioned or JSON outputs, some build structured evaluation frameworks, and some re-architect the reporting stack around evidence-linked human supervision. The common direction is the migration of radiology report generation away from unconstrained prose and toward representations in which anatomy, findings, measurements, comparison, uncertainty, and downstream computability are explicit design targets.