cs.MAOct 4, 2026
Savetemplar: agentic induction and evolution of standardized radiology reporting templates from large-scale clinical corpora
Organizations: Tsinghua University · University of Hong Kong · University of California, San Diego
Abstract
Structured radiology reporting mitigates the heterogeneity of free-text reports, yet its benefits depend on high-quality reporting templates. In practice, such templates are conventionally built through labor-intensive expert consensus and therefore vary across institutions and lag behind evolving clinical practice. Large language models (LLMs) enable automated template induction, but existing approaches remain limited: single-LLM induction is constrained by context length, and the corpus-scale method ASTAR produces a static, closed-corpus template without external grounding or downstream adaptation. To address these limitations, we propose TEMPLAR, a TEMPLate-centric Agentic framework for inducing and evolving standardized Radiology reporting templates from large-scale clinical corpora. TEMPLAR treats the template as a persistent central state maintained alongside two provenance-aware knowledge graphs, namely an anatomical graph that constrains template construction and a diagnostic graph that supports finding-to-diagnosis reasoning. Three agents operate on this state. The Induction Agent derives canonical clinical slots from anatomy-constrained Span-Triple atoms via dual-view similarity clustering; the Evolution Agent then assembles these slots into a hierarchical template and revises it under consistency constraints, external clinical evidence, and downstream structuring feedback; and the Clinical Agent applies the evolved template to report structuring, reconstruction, and diagnostic reasoning. Across four datasets, TEMPLAR outperforms ASTAR, three medical LLMs, and six general-purpose LLMs in coverage, information fidelity, and diagnostic fidelity, while achieving the highest or tied-highest LLM-rated template quality. Its fidelity advantages over ASTAR persist under cross-dataset transfer, and cumulative ablations support complementary contributions of its key components.
Figures & tables
Figure 1: Comparison of reporting-template construction paradigms. (a) Labor-intensive manual design, context-limited single-LLM induction, static corpus-scale induction with ASTAR, and adaptive agentic evolution with TEMPLAR . (b) Single-LLM templates may conflate clinical concepts and ASTAR templates retain redundant keys, whereas TEMPLAR couples an evidence-grounded, evolvable template with complementary knowledge graphs.
Figure 2: TEMPLAR overview. Anatomy-guided induction derives clinical slots, which evolve into a template through assembly, revision, external evidence grounding, and structuring feedback.
| Template Coverage | Information Fidelity | Diagnostic Fidelity | ||||||||||||
| Method | Case | Key | Avg. | R-1 | R-2 | R-L | chrF | chrF++ | BERT | Avg. | PDA | KFP | CA | Avg. |
| PETCT ( ) | ||||||||||||||
| Medical LLM baselines | ||||||||||||||
| lingshu-32b | 0.6678 | 0.7192 | 0.6935 | 0.5920 | 0.3089 | 0.3937 | 0.4916 | 0.4687 | 0.8865 | 0.5236 | 0.4939 | 0.5310 | 0.6000 | 0.5416 |
| medgemma-27b | 0.6744 | 0.7027 | 0.6886 | 0.7495 | 0.5163 | 0.5727 | 0.6214 | 0.6058 | 0.9069 | 0.6621 | 0.5633 | 0.6051 | 0.6571 | 0.6085 |
| clinfusion-32b | 0.4710 | 0.4910 | 0.4810 | 0.1419 | 0.0580 | 0.1089 | 0.0598 | 0.0556 | 0.7380 | 0.1937 | 0.5102 | 0.5320 | 0.6204 | 0.5542 |
Table 1: Performance comparison on PETCT and ViMed in template coverage, information fidelity, and diagnostic fidelity. Best in bold and runner-up in underlined . Case = case-macro coverage; Key = key-micro coverage; Avg. = average; R-1/R-2/R-L = ROUGE-1/ROUGE-2/ROUGE-L; chrF = character n-gram F-score; chrF++ = character and word n-gram F-score; BERT = BERTScore; PDA = Primary Diagnosis Accuracy; KFP = Key Finding Preservation; CA = Clinical Actionability.
| Template Coverage | Information Fidelity | Diagnostic Fidelity | ||||||||||||
| Method | Case | Key | Avg. | R-1 | R-2 | R-L | chrF | chrF++ | BERT | Avg. | PDA | KFP | CA | Avg. |
| MERLIN ( ) | ||||||||||||||
| Medical LLM baselines | ||||||||||||||
| lingshu-32b | 0.5214 | 0.5347 | 0.5281 | 0.5677 | 0.3047 | 0.3765 | 0.4965 | 0.4701 | 0.8521 | 0.5113 | 0.8662 | 0.8975 | 0.8637 | 0.8758 |
| medgemma-27b | 0.3230 | 0.3801 | 0.3516 | 0.6072 | 0.3891 | 0.3860 | 0.4705 | 0.4538 | 0.8766 | 0.5305 | 0.7137 | 0.8057 | 0.7456 | 0.7550 |
| clinfusion-32b | 0.3293 | 0.3435 | 0.3364 | 0.2769 | 0.1159 | 0.1589 | 0.2084 | 0.1947 | 0.7833 | 0.2897 | 0.5853 | 0.6650 | 0.6333 | 0.6279 |
Table 2: Performance comparison on MERLIN and BDMAP in template coverage, information fidelity, and diagnostic fidelity. Best in bold and runner-up in underlined .
Figure 3: Template quality across four datasets. Expert ratings (top) and LLM ratings (bottom) are shown for (a) PETCT, (b) ViMed, (c) MERLIN, and (d) BDMAP.
Figure 4: Cross-dataset generalization. Template coverage, information fidelity, and diagnostic fidelity under PETCT ViMed and MERLIN BDMAP transfer, without target-domain template re-induction. Open and filled circles denote ASTAR and TEMPLAR , respectively.
Figure 5: Ablation study results. Template coverage (a), information fidelity (b,c), and diagnostic fidelity (d) on PETCT, ViMed, and PETCT ViMed transfer (OOD). Starting from the baseline (M1), M2–M6 successively add the Dual-View Similarity Graph, Anatomical KG Structuring, Slot-Conditioned RAG, Template Evolution, and Diagnostic KG Reasoning.
Appendix figures & tables35 assets
Supplementary material from the paper’s appendix.
Appendix
| Category | ID | Evaluation criterion |
|---|---|---|
| Coverage and completeness | TD1 | Coverage of the essential structures and measurements required in PET/CT reports. |
| Coverage and completeness | TD2 | Explicit support for common abnormalities, including primary tumors, lymph-node metastases, and distant metastases involving the bone, liver, or lung. |
| Structure and granularity | TD3 | Hierarchical organization consistent with radiological reasoning, proceeding from findings to organ systems and key anatomical structures. |
| Structure and granularity | TD4 | Appropriate field granularity that preserves clinically relevant information without making the template unnecessarily fragmented or burdensome to complete. |
| Terminology and clarity | TD5 | Clear field names and Chinese labels that conform to commonly used clinical terminology. |
| Terminology and clarity | TD6 | Natural and usable permissible values, such as Normal , Abnormal , Uncertain , and NotAssessed . |
Table 3: Dimensions of Template Design Quality. The 13 dimensions assess the structural design, clinical safety, adaptability, and overall usability of a radiology reporting template. The criteria shown below correspond to the PET/CT instantiation.
| Category | ID | Evaluation criterion |
|---|---|---|
| Omission reduction | CU1 | The template reduces omissions of clinically important findings. |
| Reporting consistency | CU2 | The template improves consistency across PET/CT reports. |
| Workflow alignment | CU3 | The ordering of template fields is consistent with the clinician’s image-reading workflow. |
| Unassessed-state support | CU4 | The template enables unassessed structures or findings to be represented naturally using the NotAssessed state. |
| Communication efficiency | CU5 | The template improves communication efficiency between radiologists and referring clinicians. |
| Overall satisfaction | CU6 | Overall evaluator satisfaction with the template. |
Table 4: Dimensions of Clinical Utility. The 6 dimensions assess the expected value of the template in routine radiological interpretation and reporting workflows. The criteria shown below correspond to the PET/CT instantiation.
| Template Coverage | Information Fidelity | Diagnostic Fidelity | ||||||||||||
| Method | Case | Key | Avg. | R-1 | R-2 | R-L | chrF | chrF++ | BERT | Avg. | PDA | KFP | CA | Avg. |
| Medical LLM baselines | ||||||||||||||
| lingshu-32b | 0.5213 | 0.5339 | 0.5276 | 0.5171 | 0.2750 | 0.3669 | 0.3902 | 0.3742 | 0.8770 | 0.4667 | 0.4861 | 0.5172 | 0.5304 | 0.5112 |
| medgemma-27b | 0.5388 | 0.5548 | 0.5468 | 0.6382 | 0.4104 | 0.4717 | 0.4873 | 0.4762 | 0.8970 | 0.5634 | 0.5710 | 0.6328 | 0.6232 | 0.6090 |
| clinfusion-32b | 0.4894 | 0.4930 | 0.4912 | 0.3417 | 0.1849 | 0.2407 | 0.2180 | 0.2125 | 0.7995 | 0.3329 | 0.6236 | 0.6699 | 0.6873 | 0.6603 |
| Medical LLM Avg. | 0.5165 | 0.5272 | 0.5219 | 0.4990 | 0.2901 | 0.3598 | 0.3651 | 0.3543 | 0.8578 | 0.4543 | 0.5602 | 0.6066 | 0.6136 | 0.5935 |
Table 5: Performance comparison across PETCT, ViMed, MERLIN, and BDMAP in template coverage, information fidelity, and diagnostic fidelity, reported as unweighted macro-averages across the four datasets. Higher is better for all metrics. Among individual methods, best in bold and runner-up in underlined . Case = case-macro coverage; Key = key-micro coverage; Avg. = average; R-1/R-2/R-L = ROUGE-1/ROUGE-2/ROUGE-L; chrF = character n-gram F-score; chrF++ = character and word n-gram F-score; BERT = BERTScore; PDA = Primary Diagnosis Accuracy; KFP = Key Finding Preservation; CA = Clinical Actionability.
Figure 6: Dimension-wise ablation analysis on the PETCT dataset. Results are reported for (a) eight selected Design Quality dimensions and (b) five selected Clinical Utility dimensions, comparing ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. Higher scores indicate better performance on the corresponding dimension.
Figure 7: Aggregated ablation results on the PETCT dataset. Mean scores are reported over the eight selected Design Quality dimensions, five selected Clinical Utility dimensions, and all 13 selected dimensions for ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model.
Figure 8: Dimension-wise ablation analysis on the ViMed dataset. Results are reported for (a) eight selected Design Quality dimensions and (b) five selected Clinical Utility dimensions, comparing ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. Higher scores indicate better performance on the corresponding dimension.
Figure 9: Aggregated ablation results on the ViMed dataset. Mean scores are reported over the eight selected Design Quality dimensions, five selected Clinical Utility dimensions, and all 13 selected dimensions for ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model.
Figure 10: Dimension-wise ablation analysis on the MERLIN dataset. Results are reported for (a) eight selected Design Quality dimensions and (b) five selected Clinical Utility dimensions, comparing ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. Higher scores indicate better performance on the corresponding dimension.
Figure 11: Aggregated ablation results on the MERLIN dataset. Mean scores are reported over the eight selected Design Quality dimensions, five selected Clinical Utility dimensions, and all 13 selected dimensions for ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model.
Figure 12: Dimension-wise ablation analysis on the BDMAP dataset. Results are reported for (a) eight selected Design Quality dimensions and (b) five selected Clinical Utility dimensions, comparing ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model. Higher scores indicate better performance on the corresponding dimension.
Figure 13: Aggregated ablation results on the BDMAP dataset. Mean scores are reported over the eight selected Design Quality dimensions, five selected Clinical Utility dimensions, and all 13 selected dimensions for ASTAR, Ours w/o (Note + Rev.), Ours w/o Note, and the full model.
Figure 14: Design Quality evaluation on the PETCT dataset. LLM-as-judge scores of our method and baselines across the 13 design quality dimensions (DQ1–DQ13; Table 3 ) of Template Quality. Darker colors indicate higher ordinal scores.
Figure 15: Clinical Utility evaluation on the PETCT dataset. LLM-as-judge scores of our method and baselines across the 6 clinical utility dimensions (CU1–CU6; Table 4 ) of Template Quality. Darker colors indicate higher ordinal scores.
Figure 16: LLM-as-judge Template Quality evaluation on the PETCT dataset. The figure reports the mean Design Quality score, mean Clinical Utility score, and their combined Template Quality score for our method and baselines. The combined score is computed as the average of the two aspect-level mean scores.
Figure 17: Design Quality evaluation on the ViMed dataset. LLM-as-judge scores of our method and baselines across the 13 design quality dimensions (DQ1–DQ13; Table 3 ) of Template Quality. Darker colors indicate higher ordinal scores.
Figure 18: Clinical Utility evaluation on the ViMed dataset. LLM-as-judge scores of our method and baselines across the 6 clinical utility dimensions (CU1–CU6; Table 4 ) of Template Quality. Darker colors indicate higher ordinal scores.
Figure 19: LLM-as-judge Template Quality evaluation on the ViMed dataset. The figure reports the mean Design Quality score, mean Clinical Utility score, and their combined Template Quality score for our method and baselines. The combined score is computed as the average of the two aspect-level mean scores.
Figure 20: Design Quality evaluation on the MERLIN dataset. LLM-as-judge scores of our method and baselines across the 13 design quality dimensions (DQ1–DQ13; Table 3 ) of Template Quality. Darker colors indicate higher ordinal scores.
Figure 21: Clinical Utility evaluation on the MERLIN dataset. LLM-as-judge scores of our method and baselines across the 6 clinical utility dimensions (CU1–CU6; Table 4 ) of Template Quality. Darker colors indicate higher ordinal scores.
Figure 22: LLM-as-judge Template Quality evaluation on the MERLIN dataset. The figure reports the mean Design Quality score, mean Clinical Utility score, and their combined Template Quality score for our method and baselines. The combined score is computed as the average of the two aspect-level mean scores.
Figure 23: Design Quality evaluation on the BDMAP dataset. LLM-as-judge scores of our method and baselines across the 13 design quality dimensions (DQ1–DQ13; Table 3 ) of Template Quality. Darker colors indicate higher ordinal scores.
Figure 24: Clinical Utility evaluation on the BDMAP dataset. LLM-as-judge scores of our method and baselines across the 6 clinical utility dimensions (CU1–CU6; Table 4 ) of Template Quality. Darker colors indicate higher ordinal scores.
Figure 25: LLM-as-judge Template Quality evaluation on the BDMAP dataset. The figure reports the mean Design Quality score, mean Clinical Utility score, and their combined Template Quality score for our method and baselines. The combined score is computed as the average of the two aspect-level mean scores.
| Model | Finish reason | Output tokens | Outcome |
|---|---|---|---|
| seed-2.0-pro | length | 21,623 | truncated JSON |
| seed-1.8 | length | 18,836 | truncated JSON |
| seed-2.0-mini | length | 18,634 | truncated JSON |
| seed-2.0-lite | length | 11,541 | truncated JSON |
| seed-1.6 | stop | 6,982 | valid JSON in screening |
| seed-2.1-pro | – | – | no response after approximately 40 min |
Table 6: Feasibility screening of the Doubao Seed series on a real induction prompt. No request-level output-token limit was specified.
| Dataset | Peak input | Peak output | Sum of peaks |
|---|---|---|---|
| PET/CT | 24,872 | 13,332 | 38,204 |
| ViMed | 29,387 | 8,964 | 38,351 |
| BDMAP | 39,461 | 3,558 | 43,019 |
| MERLIN | 30,039 | 13,585 | 43,624 |
Table 7: Observed peak input and output lengths for induction, measured with the Qwen tokenizer on completed runs. The final column sums the two peaks, which need not occur in the same request.
| Cause | Count | Structural signature |
|---|---|---|
| Truncation | 78 | 1–59 missing } and up to 14 missing ] |
| Generation restart | 17 | a second JSON fence or a second top-level desc |
| Excess closing delimiter | 17 | an otherwise complete object followed by extra delimiters |
| Balanced but invalid syntax | 3 | e.g., an unambiguous trailing comma |
Table 8: Structural causes of the 115 outputs rejected by strict JSON parsing. Bracket accounting ignores delimiters that occur inside strings.
| Dataset | Examples | Categorical | Numerical |
|---|---|---|---|
| PETCT | 12 | 11 | 1 |
| ViMed | 12 | 11 | 1 |
| MERLIN | 14 | 13 | 1 |
| BDMAP | 12 | 8 | 4 |
| Total | 50 | 43 | 7 |
Table 9: Distribution of representative examples from the final templates presented in this appendix.
Figure 26: Visualization of the PETCT anatomical knowledge graph. The displayed subgraph organizes representative anatomical entities under their parent anatomical region and connects them to structural, measurement, attribute, finding, and status concepts. Node and edge colors distinguish entity and relation classes, respectively, as defined in the legend. Faint unlabeled branches indicate additional observed triples included to preserve the local graph structure.
Figure 27: Concept-level diagnostic knowledge graph for PETCT. Representative canonical imaging-evidence concepts are connected to diagnostic targets through the extracted diagnostic relation types. Node area is proportional to the supporting raw-triple count, while arrow labels specify the relation between each evidence concept and diagnostic target. Gray dashed links indicate visual grouping and are not knowledge-graph edges.
Figure 28: Visualization of the ViMed anatomical knowledge graph. The displayed subgraph organizes representative anatomical entities under their parent anatomical region and connects them to structural, measurement, attribute, finding, and status concepts. Node and edge colors distinguish entity and relation classes, respectively, as defined in the legend. Faint unlabeled branches indicate additional observed triples included to preserve the local graph structure.
Figure 29: Concept-level diagnostic knowledge graph for ViMed. The visualization connects representative canonical imaging-evidence concepts to diagnostic targets through the extracted diagnostic relation types. Node area is proportional to the supporting raw-triple count, while arrow labels specify the relation between each evidence concept and diagnostic target. Gray dashed links indicate visual grouping and are not knowledge-graph edges.
Figure 30: Visualization of the MERLIN anatomical knowledge graph. The displayed subgraph organizes representative anatomical entities under their parent anatomical region and connects them to structural, measurement, attribute, finding, and status concepts. Node and edge colors distinguish entity and relation classes, respectively, as defined in the legend. Faint unlabeled branches indicate additional observed triples included to preserve the local graph structure.
Figure 31: Concept-level clinical-significance knowledge graph for MERLIN. Representative canonical report-evidence concepts are connected to clinical-significance outcomes that characterize the reported abnormality and its potential need for clinical action. Node area is proportional to the supporting raw-triple count, while arrow labels specify the extracted evidence-to-outcome relation. Gray dashed links indicate visual grouping and are not knowledge-graph edges.
Figure 32: Visualization of the BDMAP anatomical knowledge graph. The displayed subgraph organizes representative anatomical entities under their parent anatomical region and connects them to structural, measurement, attribute, finding, and status concepts. Node and edge colors distinguish entity and relation classes, respectively, as defined in the legend. Faint unlabeled branches indicate additional observed triples included to preserve the local graph structure.
Figure 33: Concept-level diagnostic knowledge graph for BDMAP. Representative canonical imaging-evidence concepts are connected to diagnostic and clinical-status targets through the extracted diagnostic relation types. Node area is proportional to the supporting raw-triple count, while arrow labels specify the relation between each evidence concept and target. Gray dashed links indicate visual grouping and are not knowledge-graph edges.