ASTAR: Automated induction of STAndardized radiology Reporting templates from large-scale clinical free-text corpora

# ASTAR: Automated induction of STAndardized radiology Reporting templates from large-scale clinical free-text corpora
Source: [https://arxiv.org/html/2608.20369](https://arxiv.org/html/2608.20369)
11institutetext:Tsinghua University, Beijing, China22institutetext:Sichuan University, Chengdu, China33institutetext:University of California San Diego, San Diego, USA44institutetext:Southeast University, Nanjing, China
44email:qiyuantian@tsinghua\.edu\.cnMingxuan Liu†Yifei ChenJuncheng ZhuKasidit AnmahapongYiming HuangYuan ZhangHongjia YangYi LiaoGang NingHaibo QuQiyuan Tian∗

###### Abstract

Structured reporting converts free\-text radiology narratives into queryable data keys, facilitating cohort assembly, longitudinal tracking, and training label generation for medical AI\. The prevailing paradigm follows a two\-stage pipeline: \(1\) constructing a reporting template, \(2\) extracting information to populate it\. While the extraction stage has benefited from advances in large language models \(LLMs\), template construction remains a manual bottleneck relying on labor\-intensive expert consensus that is static, difficult to scale, and may fail to capture real\-world reporting diversity\. We address this limitation withASTAR, an LLM\-based framework for Automated induction of STAndardized radiology Reporting templates from large\-scale clinical free\-text corpora\. Extensive experiments on 4,215 fetal brain MRI reports from multiple centers demonstrate that theASTAR\-induced template surpasses two expert\-curated templates across template coverage, information fidelity, diagnostic fidelity, and expert\-rated usability, reducing template development from weeks of committee deliberation to hours of automated processing\. Code:[https://github\.com/birthlab/ASTAR](https://github.com/birthlab/ASTAR)

††footnotetext:†Equal contribution\.∗Corresponding author\.## 1Introduction

Radiology departments worldwide generate clinical reports at an ever\-growing scale\. In the United States alone, over 16\.1 million CT examinations were reported during a nine\-month period in 2020\[Davenport2021\], and imaging utilization across all modalities is projected to rise 17–27% by 2055\[Christensen2025\]\. Each examination produces a detailed free\-text narrative synthesizing categorical descriptors, quantitative measurements, and diagnostic impressions, thereby forming one of the richest data repositories in modern healthcare\[nowak\_privacy\-ensuring\_2025,chen2025mmlnbmultimodallearningneuroblastoma\]\. However, the utility of these reports is constrained because findings remain archived as unstructured text with wide stylistic, terminological, and structural variations across radiologists, institutions, and languages\[liu\_fetalextract\-llm\_nodate,nowak\_transformer\-based\_2023\]\. Such heterogeneity hinders systematic case retrieval, cohort assembly, longitudinal tracking, and training\-label generation for medical AI\[chen\_burextract\-llama\_2024,nowak\_privacy\-ensuring\_2025\]\. Radiology societies therefore increasingly advocate for structured reporting\[sofia\_standardised\_2024,busch\_large\_2024\], converting narratives into queryable keys that support rapid cohort identification and high\-quality ground\-truth generation\[bai\_chest\-omdl\_nodate,bai2026exactexplainableanomalyawarevision\]\.

Nevertheless, achieving these benefits at scale requires automated tools, since manual structuring of the massive and continuously growing volume of narrative reports is highly impractical\. Therefore, recent studies increasingly employ natural language processing \(NLP\) and large language models \(LLMs\) for automated information extraction\[fytas\-etal\-2024\-rule,Zhang2026ContentGeneration\]\. Specifically, the prevailing paradigm follows a two\-stage pipeline: \(1\) a reporting template \(i\.e\., a target template defining the set of keys and their permissible values\) is constructed; and \(2\) relevant information is extracted from free\-text reports to populate this template\. For example, BURExtract\-Llama\[chen\_burextract\-llama\_2024\]extracts clinical concepts from breast ultrasound reports with an average F1 score of 84\.6%\.

![Refer to caption](https://arxiv.org/html/2608.20369v1/x1.png)Figure 1:\(A\) Traditional template construction pipeline\. \(B\) Our automated template induction pipeline \(ASTAR\)\.However, while the extraction stage has benefited substantially from recent advances in NLP and LLMs, the template construction stage, as an upstream prerequisite for all extraction methods, remains a manual bottleneck, since existing templates are typically produced through labor\-intensive expert consensus \(Fig\.[1](https://arxiv.org/html/2608.20369#S1.F1)A\)\. For example, the ESPR template\[sofia\_standardised\_2024\]was derived from multidisciplinary panel discussions, and the JSON template used by FetalExtract\-LLM\[liu\_fetalextract\-llm\_nodate\]was compiled from ESPR guidelines\[sofia\_standardised\_2024\], MeSH\[lipscomb2000medical\], and a medical dictionary\. Beyond this scalability limitation, expert\-designed templates are inherently static and may fail to capture the full spectrum of linguistic diversity and evolving clinical descriptions present in massive historical report corpora\[busch\_large\_2024\]\. As a result, rare but clinically meaningful findings risk being absent from the predefined template, while institution\-specific reporting conventions cannot be accommodated by a one\-size\-fits\-all structure\.

To the best of our knowledge, no automated method exists for standardized radiology reporting templates construction\. To fill this gap, we proposeASTAR\(Automated induction of STAndardized radiology Reporting templates\), an LLM\-based framework \(Fig\.[1](https://arxiv.org/html/2608.20369#S1.F1)B\) that automatically mines unified reporting structures from large\-scale radiology report corpora\. We demonstrate its utility on 4,215 fetal brain MRI reports from three centers, where theASTAR\-induced template achieved stable information retention with progressive consolidation of long\-tail keys as corpus size increases, offering a scalable, data\-driven complement to labor\-intensive expert consensus\.

## 2Method

### 2\.1Overview ofASTAR

Given a corpus ofNNde\-identified free\-text radiology reports𝒟=\{r1,…,rN\}\\mathcal\{D\}=\\\{r\_\{1\},\\dots,r\_\{N\}\\\}, the goal ofASTARis to automatically induce a hierarchical reporting template𝒯=\(𝒱,ℰ,\{fv\}v∈𝒱leaf\)\\mathcal\{T\}=\(\\mathcal\{V\},\\,\\mathcal\{E\},\\,\\\{f\_\{v\}\\\}\_\{v\\in\\mathcal\{V\}\_\{\\text\{leaf\}\}\}\), where internal nodes represent anatomical or semantic groupings and each leaf nodev∈𝒱leafv\\in\\mathcal\{V\}\_\{\\text\{leaf\}\}carries a field descriptorfv=\(kv,τv,Ωv\)f\_\{v\}=\(k\_\{v\},\\tau\_\{v\},\\Omega\_\{v\}\)specifying a standardized field namekvk\_\{v\}, a value typeτv∈\{categorical,numerical,free\-text\}\\tau\_\{v\}\\in\\\{\\texttt\{categorical\},\\allowbreak\\texttt\{numerical\},\\allowbreak\\texttt\{free\\mbox\{\-\}text\}\\\}, and a type\-dependent value specificationΩv\\Omega\_\{v\}\.ASTARobtains𝒯\\mathcal\{T\}through three stages \(Fig\.[2](https://arxiv.org/html/2608.20369#S2.F2)\):𝒟→IΦ→IIT1→III𝒯∗\\mathcal\{D\}\\xrightarrow\{\\textbf\{I\}\}\\Phi\\xrightarrow\{\\textbf\{II\}\}T\_\{1\}\\xrightarrow\{\\textbf\{III\}\}\\mathcal\{T\}^\{\*\}, where Stage I builds a de\-duplicated concept slots libraryΦ=\{ϕ1,…,ϕL\}\\Phi=\\\{\\phi\_\{1\},\\dots,\\phi\_\{L\}\\\}with eachϕj=\(kj,τj,Ωj\)\\phi\_\{j\}=\(k\_\{j\},\\tau\_\{j\},\\Omega\_\{j\}\)sharing the same structure asfvf\_\{v\}; Stage II organizesΦ\\Phiinto a hierarchical tree, mapping each surviving slot to exactly one leaf \(fv←ϕjf\_\{v\}\\\!\\leftarrow\\\!\\phi\_\{j\}\); and Stage III refines the tree into the final template𝒯∗\\mathcal\{T\}^\{\*\}\.

### 2\.2Stage I: Concept Slots Library Construction

In stage I, raw narratives are converted into a de\-duplicated library of canonical clinical concept slots in two steps\. \(1\)Span Extraction\.An LLM \(Qwen\-Max\[qwen3\]\) in JSON\-mode segments each reportrir\_\{i\}into spans𝒮i\\mathcal\{S\}\_\{i\}, each represented as a 7\-tuple\(astruct,asub,aattr,atype,aval,arole,asec\)\(a\_\{\\text\{struct\}\},\\,a\_\{\\text\{sub\}\},\\,a\_\{\\text\{attr\}\},\\,a\_\{\\text\{type\}\},\\,a\_\{\\text\{val\}\},\\,a\_\{\\text\{role\}\},\\,a\_\{\\text\{sec\}\}\)encoding anatomical location \(structure / substructure\), attribute identity \(name / type\), a canonicalized value \(numerics abstracted as⟨NUM⟩\\langle\\texttt\{NUM\}\\rangle\), clinical role \(finding​∣measurement∣​diagnosis\\texttt\{finding\}\\mid\\texttt\{measurement\}\\mid\\texttt\{diagnosis\}\), and report section \(description∣impression\\texttt\{description\}\\mid\\texttt\{impression\}\)\. Spans from all reports are pooled and de\-duplicated by the composite key\(aval,arole\)\(a\_\{\\text\{val\}\},a\_\{\\text\{role\}\}\)into a global span set𝒮\\mathcal\{S\}\. \(2\)Clustering and Slot Induction\.Each span is embedded via Qwen3\-Embedding\-8B\[qwen3\]and grouped byKK\-means intoCCclusters, with the optimal number of clustersC∗C^\{\*\}chosen by maximizing the silhouette score\[ROUSSEEUW198753\]overC∈\[50,300\]C\\\!\\in\\\!\[50,300\]\. Because raw clusters may still mix heterogeneous concepts \(e\.g\.,ventricular widthandventricular morphology\), the LLM sub\-partitions each cluster into semantically coherent subgroups and induces a canonical slotϕ=\(k,τ,Ω\)\\phi=\(k,\\,\\tau,\\,\\Omega\)per subgroup, wherekkis a standardized key name,τ∈\{categorical,numerical,free\-text\}\\tau\\\!\\in\\\!\\\{\\texttt\{categorical\},\\texttt\{numerical\},\\texttt\{free\-text\}\\\}the value type, andΩ\\Omegathe permissible value set\. Slots sharing identical standardized key names \(k\) are then globally merged by the LLM, yielding the unified concept slots libraryΦ=\{ϕ1,…,ϕL\}\\Phi=\\\{\\phi\_\{1\},\\dots,\\phi\_\{L\}\\\}\.

![Refer to caption](https://arxiv.org/html/2608.20369v1/x2.png)Figure 2:Overview ofASTAR\. \(1\) Stage I: Concept Slots Library Construction\. \(2\) Stage II: Hierarchical Template Tree Construction\. \(3\) Stage III: Template Refinement\.
### 2\.3Stage II: Hierarchical Template Tree Construction

In Stage II, given the slots libraryΦ\\Phi, the flat set of slots is organized into a clinically meaningful hierarchy through a two\-round LLM\-driven process\.

Round 1: Template Construction and Initial Assignment\.An LLM \(Qwen\-Max\) first generates an initial Template𝒯0\\mathcal\{T\}\_\{0\}from a representative subset of reports, with internal nodes organized by major anatomical regions\. Each slotϕj∈Φ\\phi\_\{j\}\\in\\Phiis then presented to the LLM together with its descriptive statistics and the current template\. The LLM decides toattachit under an existing node,proposea new branch, ordropit\. Proposed branches are incorporated into the template incrementally\.Round 2: Global Path Refinement\.Because Round 1 slots are assigned against an evolving template, early slots never see branches proposed later\. A second pass therefore re\-evaluates all surviving slots in batches on the now\-complete tree, taking each slot’s Round 1 path as reference, and outputs a revised path, yielding𝒯1\\mathcal\{T\}\_\{1\}\.

### 2\.4Stage III: Template Refinement

The final stage performs two refinement operations on𝒯1\\mathcal\{T\}\_\{1\}to ensure global consistency and output the final template𝒯∗\\mathcal\{T\}^\{\*\}\. \(1\)Intra\-node Semantic Reorganization\.Within each node, the LLM consolidates synonymous permissible values \(e\.g\., \[“normal”,“no abnormality”,“unremarkable”\]→\\to\[“Normal”\]\) and harmonizes value type assignments across sibling keys, reducing value\-level redundancy while preserving clinical granularity\. \(2\)Global Concept Harmonization\.A tree\-wide pass identifies cross\-node redundancies, i\.e\., keys whose names differ across subtrees but refer to the same concept, and standardizes terminology throughout the tree \(e\.g\., unifying“cerebral hemisphere”vs\.“brain hemisphere”across subtrees\)\.

### 2\.5Evaluation Protocol for Radiology Reporting Templates

Because there is no widely used protocol for evaluating radiology reporting templates, we introduce a three\-level protocol to fill this gap: \(1\)Template qualityassesses whether the template is comprehensive and well\-structured; \(2\)structuring fidelitymeasures whether populating the template preserves the information in the original report; and \(3\)radiologist judgmentcaptures whether clinicians find the template usable and clinically valuable\. Specifically,

\(1\)Template quality, computed on a held\-out test dataset, quantifies how completely𝒯∗\\mathcal\{T\}^\{\*\}captures the clinical concepts present in unseen reports, measured by*case\-level*\(macro\-averaged over reports\) and*key\-level*\(micro\-averaged over all keys\) coverage\. For each test reportrir\_\{i\}, concept keysKiK\_\{i\}are extracted by concatenating entity and attribute keys of each span\. An LLM \(Qwen\-Max\) judges whether each key has a semantic match in the template key setℱ​\(𝒯∗\)\\mathcal\{F\(\\mathcal\{T\}^\{\*\}\)\}\. The two metrics are defined as:

Covcase=1\|𝒟test\|​∑i∑k∈Kimatch​\(k,ℱ\)\|Ki\|,Covkey=∑i∑k∈Kimatch​\(k,ℱ\)∑i\|Ki\|\.\\mathrm\{Cov\}\_\{\\text\{case\}\}=\\frac\{1\}\{\|\\mathcal\{D\}\_\{\\text\{test\}\}\|\}\\sum\_\{i\}\\frac\{\\sum\_\{k\\in K\_\{i\}\}\\mathrm\{match\}\(k,\\mathcal\{F\}\)\}\{\|K\_\{i\}\|\},\\quad\\mathrm\{Cov\}\_\{\\text\{key\}\}=\\frac\{\\sum\_\{i\}\\sum\_\{k\\in K\_\{i\}\}\\mathrm\{match\}\(k,\\mathcal\{F\}\)\}\{\\sum\_\{i\}\|K\_\{i\}\|\}\.\(1\)
\(2\)Structuring fidelity, also computed on the held\-out test dataset, measures whether populating the template preserves the information in the original report, assessed via*information fidelity*and*diagnostic fidelity*\. Specifically, \(a\)*information fidelity*, evaluated through a four\-step extract–reconstruct round trip, quantifies textual information preservation\.\(i\)An LLM \(Gemini\-2\.5\-Pro\-Preview\[comanici2025gemini\]\), chosen independently of the template\-construction LLM to prevent extraction bias, populates the template from each report’simaging description;\(ii\)extracted key–value pairs are mapped into the full hierarchy and re\-flattened for cross\-template consistency;\(iii\)another LLM \(Qwen\-Max\) reconstructs a free\-text report solely from the structured report, guided by a fixed style\-reference text;\(iv\)the reconstructionr^i\\hat\{r\}\_\{i\}is compared with the originalrir\_\{i\}using ROUGE\-1/2/L \(F1\)\[lin\-2004\-rouge\], chrF\[popovic2015chrf\], BERTScoreR\(F1\) \(with RoBERTa\-wwm\-ext\[cui2021pre\]\), and BERTScoreM\(F1\)\[zhang2020bertscore\]\(with MacBERT\[cui2020revisiting\]\)\. \(b\)*Diagnostic Fidelity*, evaluated through an infer\-then\-judge protocol, tests whether the structured report retains sufficient clinical information for diagnosis\. Qwen\-Max receives only the structured output, infers a diagnosisdistructd\_\{i\}^\{\\text\{struct\}\}under strict information closure, and an LLM\-as\-Judge protocol \(with Qwen\-Max\) is used to compare it against the ground\-truth impressiondigtd\_\{i\}^\{\\text\{gt\}\}\(i\.e\., thediagnostic impressionin the original free\-text report\) using three metrics \(all normalized to\[0,1\]\[0,1\]\):*Primary Diagnosis Accuracy*\(PDA; 0–5 integer, macro correctness\),*Key Finding Preservation*\(KFP; 0–1 continuous, abnormal\-finding retention\), and*Clinical Actionability*\(CA; 0–5 integer, decision\-support sufficiency\)\.

\(3\)Radiologist judgmentdirectly evaluates the template via structured clinician ratings\. Two radiologists from different countries, ensuring cross\-national diversity in clinical conventions and mitigating institution\-specific bias, rated all three templates on 28 Likert\-scale items \(1–5\) across three dimensions\. \(a\)*Template Structure*\(12 items\) evaluates the schema design across six sub\-dimensions: coverage & completeness \(comp,abno\), structure & granularity \(logi,gran\) terminology & clarity \(name,enum\), safety & error prevention \(noms,nofn,unce\), adaptability \(adap\), and deployment readiness \(adop,sign\); \(b\)*System Usability*\(10 items\) follows the standard scale\[brooke1996sus\], comprisingfreq\(willingness to use\),comp\(complexity†\),ease\(ease of use\),tech\(technical support needed†\),inte\(integration\),inco\(inconsistency†\),lear\(learnability\),cumb\(cumbersomeness†\),conf\(confidence\), andprel\(prior learning needed†\), where†\\daggermarks negatively worded items reverse\-scored as6−xraw6\{\-\}x\_\{\\text\{raw\}\}; \(c\)*Clinical Impact*\(6 items\) assesses downstream utility:noms\(omission reduction\),cons\(consistency\),orde\(workflow alignment\),naex\(not\-assessed expressibility\),comm\(communication efficiency\), andovr\(overall satisfaction\)\.

Table 1:Template quality \(coverage\) and structuring fidelity \(\(information fidelity & diagnostic fidelity\)\) across templates evaluated on ID \(n=100n\{=\}100\) and OoD \(n=115n\{=\}115\) test sets\. Bold indicates best results\. The three templates differ in key count \(ESPR: 159, FetalExtract: 51,ASTAR: 88\), and the results suggest that key relevance and organization, rather than key count alone, contribute to the observed differences\.![Refer to caption](https://arxiv.org/html/2608.20369v1/x3.png)Figure 3:Two radiologists from different countries rated all templates on 28 Likert\-scale items \(11–55, higher is better\)\.\(A\)Radiologist 1;\(B\)Radiologist 2\.Top row:per\-item radar charts for Template Structure \(12 items\), System Usability \(10 items\), and Clinical Impact \(6 items\); numbers on each axis denote the maximum score among the three templates for that item\.Bottom row:dimension\-level means; dashed lines show cross\-radiologist averages\.

## 3Experiments and Results

### 3\.1Data Collection

A total of 4,100 de\-identified fetal brain MRI free\-text reports were retrospectively collected from the \[Anonymous\] hospital, covering examinations performed between October 2015 and December 2024 on one 3\.0 T system \(Siemens MAGNETOM Skyra\) and two 1\.5 T systems \(Philips Achieva and United Imaging uMR 570\)\. Each report comprises a narrativeimaging descriptionsection and adiagnostic impressionsection\. Of these, 4,000 reports were used for template construction withASTARand 100 were held out as an in\-distribution \(ID\) test dataset\. To assess cross\-institutional generalizability, an additional out\-of\-distribution \(OoD\) test dataset of 115 de\-identified fetal brain MRI reports was collected from \[Anonymous\] hospital and \[Anonymous\] hospital\.

### 3\.2Comparison with Expert\-Designed Templates

The ASTAR\-induced template was compared against two expert\-designed templates\. \(1\)ESPR Template\[sofia\_standardised\_2024\]: a standardized fetal MRI reporting template released in 2024 by the Fetal Task Force of the European Society of Paediatric Radiology\. \(2\)FetalExtract Template\[liu\_fetalextract\-llm\_nodate\]: a JSON template compiled from ESPR guidelines, MeSH, and a medical dictionary, used for structured information extraction with FetalExtract\-LLM\.

Across all metrics,ASTARconsistently outperforms both expert\-designed templates on both ID and OoD datasets, achieving higher coverage with fewer keys, superior information and diagnostic fidelity, and the highest expert\-rated usability \(Table[1](https://arxiv.org/html/2608.20369#S2.T1), Fig\.[3](https://arxiv.org/html/2608.20369#S2.F3)\)\. \(1\)Template Quality:On the ID test dataset,ASTARachieves case\-/key\-level coverage of 81\.34%/82\.42%, outperforming ESPR and FetalExtract\. This advantage persists on the OoD set, though the gap narrows due to expected domain shift\. \(2\)Structuring Fidelity:ASTARconsistently achieves the highest scores across all six metrics for*information fidelity*on both datasets\. Specifically, on the ID dataset,ASTARattained ROUGE\-L0\.694±0\.1000\.694\\pm 0\.100, substantially outperforming ESPR and FetalExtract; BERTScoreRreached0\.895±0\.0160\.895\\pm 0\.016vs\.0\.892±0\.0160\.892\\pm 0\.016and0\.882±0\.0210\.882\\pm 0\.021\. On the OoD test dataset,ASTARmaintains a clear lead \(ROUGE\-L0\.503±0\.1920\.503\\pm 0\.192vs\.0\.394±0\.1750\.394\\pm 0\.175and0\.305±0\.2190\.305\\pm 0\.219\), with larger standard deviations reflecting greater cross\-institutional stylistic heterogeneity\. For*diagnostic fidelity*,ASTARachieves the highest scores across all three dimensions on the ID set: PDA0\.953±0\.0900\.953\\pm 0\.090, KFP0\.958±0\.0930\.958\\pm 0\.093, and CA0\.953±0\.0850\.953\\pm 0\.085, outperforming ESPR \(PDA/KFP/CA:0\.8790\.879/0\.8820\.882/0\.8760\.876\) and FetalExtract \(0\.8580\.858/0\.8440\.844/0\.8580\.858\)\. On the OoD set,ASTARretains its advantage \(PDA0\.8470\.847, KFP0\.8110\.811, CA0\.8470\.847\), with consistently high KFP \(\>\>0\.81\) confirming that the automatically induced template preserves critical abnormal findings \(e\.g\., ventriculomegaly, posterior fossa anomalies\) essential for clinical decision\-making\. \(3\)Radiologist Judgment:ASTARreceives the highest scores across*Template Structure*4\.33/5\.00 \(vs\. ESPR 3\.88, FetalExtract 2\.88\),*System Usability*3\.90 \(vs\. 3\.70, 3\.20\), and*Clinical Impact*4\.33 \(vs\. 4\.17, 3\.42\)\.

![Refer to caption](https://arxiv.org/html/2608.20369v1/x4.png)Figure 4:Scaling behavior of ASTAR\-induced templates on the ID test set \(n=100n\{=\}100\)\.\(A\)Case\-level coverage \(left axis\), key\-level coverage \(orange, left axis\) and key coverage tail ratio \(gold, right axis\)\.\(B\)ROUGE\-L distributions \(left axis\), mean ROUGE\-L \(blue, left axis\) and report\-level tail ratio \(red dashed, right axis\)\.
### 3\.3Scaling Behavior ofASTAR

To investigate how corpus size affects template quality, we constructedASTARtemplates from subsets ofn=32n\{=\}32ton=4000n\{=\}4000reports in the training dataset and evaluated each on the ID test dataset \(Fig\.[4](https://arxiv.org/html/2608.20369#S3.F4)\)\. Key\-level coverage remains consistently high \(0\.767–0\.859\), while the tail ratio \(case\-level coverage < 0\.01\) decreases from 39\.7% atn=32n\{=\}32to 14\.5% atn=2048n\{=\}2048before stabilizing, confirming thatASTARprogressively consolidates rare\-but\-valid concepts into coherent keys \(Fig\.[4](https://arxiv.org/html/2608.20369#S3.F4)A\)\. Additionally, mean ROUGE\-L remains stable across all corpus sizes \(0\.676–0\.696\), while the report\-level tail ratio \(case\-level ROUGE\-L < 0\.5\) drops from 9\.8–10\.5% atn≤64n\\leq 64to 3\.1% atn=4000n\{=\}4000, indicating that additional reports primarily improve coverage of structurally atypical or linguistically rare cases rather than boosting average fidelity \(Fig\.[4](https://arxiv.org/html/2608.20369#S3.F4)B\)\.

## 4Conclusion

We presentedASTAR, the first LLM\-based framework that automatically induces standardized radiology reporting templates from large\-scale free\-text corpora\. Evaluated on 4,215 multi\-center fetal brain MRI reports, the ASTAR\-induced template outperforms two expert\-designed baseline templates in template quality, structuring fidelity, and radiologist\-rated usability, reducing template development from weeks of expert consensus to hours of automated processing\.

## References