PaperlayerView exact revision on arXiv

Source-linked AI summary

STRIVE: Multi-Agent Structured Temporal Reasoning with Integrated Verification for Longitudinal Radiology Report Generation

Junyeong Maeng, Eunsong Kang, Heung-Il Suk

arXiv:2608.24237v1cs.AI

TL;DR

Longitudinal radiology report generation must capture both current findings and their changes over time, but existing approaches provide limited explicit, verifiable temporal reasoning. STRIVE uses specialized agents and integrated verification, improving report quality, clinical efficacy, and temporal agreement on Longitudinal-MIMIC while more than doubling LCC over the strongest baseline.

  • Problem

    Existing LRRG methods rely on implicit shared representations, while progression-specific properties are assessed only indirectly rather than through direct temporal evaluation.

  • Method

    STRIVE assigns diagnosis, attribute estimation, and temporal reasoning to specialized agents, using progression-aware training, consistency reconciliation, and report-level validation.

  • Results

    On Longitudinal-MIMIC, STRIVE improves linguistic fluency, diagnostic performance, and progression-state agreement, achieving LCC-C 0.394 versus 0.193 for the strongest baseline.

  • Takeaways & Limitations

    The larger clinical-efficacy gains than NLG gains suggest that STRIVE’s improvements primarily reflect clinical correctness rather than surface-level similarity.

  • Takeaways & Limitations

    The evaluation remains limited because overlap metrics do not distinguish clinically decisive terms from stylistic words, while CE measures current findings rather than temporal direction.

Abstract

from arXiv · show

Longitudinal radiology report generation (LRRG) requires identifying both current findings and their changes relative to a prior study. Existing methods jointly model diagnosis, attribute estimation, temporal comparison, and language generation within implicit representations, which can cause task interference, obscure the evidence underlying each decision, and limit error traceability. They also model progression states as independent labels, ignoring their ordered structure and thus treating missed changes and direction reversals equally. We present STRIVE, Multi-Agent Structured Temporal Reasoning with Integrated Verification for LRRG, which decomposes clinical reasoning into specialized Diagnosis, Attribute, and Temporal Change Agents that produce explicit intermediate evidence. In particular, the Temporal Change Agent is further post-trained using Progression-Aware GRPO, a verifiable, shaped reward that assigns partial credit to direction-preserving errors while scoring direction reversals lowest. STRIVE performs verification at two stages: a deterministic Consistency Gate reconciles the agent outputs before report generation, and a Validation Agent checks whether the generated report is supported by the aggregated clinical evidence. On Longitudinal-MIMIC, STRIVE attains the best clinical efficacy among recent methods and more than doubles Longitudinal Change Concordance (LCC), a measure of temporal agreement with the reference report, over the strongest baseline.

Introduction

STRIVE addresses limitations in longitudinal radiology report generation by separating clinical reasoning into specialized agents with explicit evidence, modeling ordered progression states, and verifying consistency and report support. It is evaluated on Longitudinal-MIMIC for linguistic fluency, diagnostic performance, and progression-state agreement measured by Longitudinal Change Concordance (LCC).

  • Background: Radiology report generation requires identifying fine-grained clinical information, including disease presence, severity, and location, and expressing it faithfully in natural language.Recent vision-language and large language models have improved generated-report fluency and clinical accuracy.
  • Limitations: Existing methods implicitly couple diagnosis, attribute estimation, temporal reasoning, and report generation, causing task interference, obscured evidence, and difficult error localization.Clinical objectives require discrete, constrained decisions, whereas report generation favors a smooth semantic space.
  • Limitations: Existing objectives model progression states independently, overlooking directional and ordered relationships among changes such as new, resolved, increased, stable, and decreased.This limitation motivates progression-aware modeling rather than treating all state errors equivalently.
  • STRIVE: STRIVE decomposes diagnosis, clinical attribute estimation, and temporal progression modeling into specialized agents that produce explicit intermediate evidence.The framework is designed to enable task-specific optimization while making clinical evidence explicit.
  • STRIVE: STRIVE introduces Progression-Aware GRPO, a shaped verifiable reward that captures directional and graded relationships among longitudinal progression states.It also uses a Consistency Gate to resolve agent conflicts and a Validation Agent to verify that the final report is supported by the committed clinical state.
  • Evaluation: STRIVE outperforms existing methods on Longitudinal-MIMIC in linguistic fluency, diagnostic performance, and progression-state agreement measured by Longitudinal Change Concordance (LCC).The introduction identifies LCC as the measure of progression-state agreement.

Related Work

Related work progresses from single-study radiology report generation toward longitudinal methods that compare prior and current studies. However, existing LRRG systems largely rely on shared implicit representations, limiting error traceability and verification across heterogeneous reasoning tasks.

  • Single-study and longitudinal report generation: Conventional RRG recognizes clinical findings such as disease presence, severity, and location from a single study, without directly modeling cross-time disease changes.LRRG extends RRG by incorporating historical studies and modeling temporal evidence.
  • Longitudinal modeling strategies: Existing LRRG methods differ in how they relate prior and current studies, using fusion-based unified representations or alignment-based cross-time relationships.PriorRG fuses prior and current visual features, while alignment-based methods establish relationships between local visual regions.
  • Limitations of existing LRRG: Despite architectural diversity, existing LRRG methods jointly encode disease recognition, attribute estimation, and temporal comparison within shared implicit representations.This joint modeling causes interference among heterogeneous reasoning tasks and makes errors difficult to trace and verify.
  • Multi-agent report generation: Recent RRG studies adopt multi-agent formulations that structure image interpretation and report generation into specialized stages based on clinical workflows.CogRad assigns triage, investigation, writing, and verification to specialized agents, while RadAgents divides interpretation by anatomical region.

Method

STRIVE decomposes longitudinal report generation into specialized clinical decisions with explicit intermediate evidence, then applies progression-aware optimization and staged verification before producing the final report.

  • Task decomposition: STRIVE factorizes longitudinal report generation into diagnosis, attribute estimation, and temporal-change decisions handled by role-specialized agents.The Diagnosis Agent detects current findings, the Attribute Agent characterizes positive findings, and the Temporal Change Agent identifies changes since the prior study.
  • Temporal reasoning: The Temporal Change Agent predicts six ordered change states for each disease-related finding using prior reports, diagnosis states, probabilities, and the inter-study interval.Its labels are new, increased, stable, decreased, resolved, and none.
  • Progression-aware optimization: Progression-Aware GRPO uses a programmatic, multi-level reward that gives partial credit to direction-preserving errors and penalizes opposite-direction predictions most strongly.The reward evaluates detection, coarse clinical direction, and fine-grained change-state agreement within an RLVR pipeline.
  • Verification and evidence aggregation: A deterministic Consistency Gate reconciles incompatible diagnosis and temporal-change outputs before aggregating them into a Structured Clinical State.The SCS preserves disease-wise evidence for report generation and supports tracing errors to their corresponding clinical decisions.
  • Report generation and validation: Report generation conditions a frozen Writer on the Structured Clinical State, base image-generated draft, retrieved exemplars, and clinical context, followed by Validation Agent edits.Validation checks draft diagnoses and temporal descriptions against the SCS, adding missing supported findings and correcting or removing inconsistent content under stated evidence conditions.

Experiments

STRIVE is evaluated on Longitudinal-MIMIC using standard language, clinical efficacy, ReXrank, and longitudinal change metrics. It achieves strong report quality and substantially improves temporal-change agreement, with ablations attributing longitudinal correctness primarily to the Temporal Change Agent.

  • Dataset: Evaluation uses Longitudinal-MIMIC, pairing each current study with the patient’s most recent prior study and report.The reference report describes current findings and their changes relative to the prior study.
  • Overall results: STRIVE achieves the best performance on all NLG metrics except ROUGE-L and on all three clinical efficacy metrics.Larger clinical-efficacy gains than NLG gains suggest improvement in clinical correctness rather than surface similarity.
  • ReXrank results: STRIVE leads six of seven ReXrank metrics: 1/RadCliQ-v1 (1.173), BLEU (0.272), SembScore (0.467), RadGraph-F1 (0.270), RaTEScore (0.582), and GREEN (0.342).It ranks second on BERTScore.
  • Longitudinal correctness: 0.394 and 0.283 are STRIVE’s LCC-C and LCC-F scores, more than twice the strongest baseline scores of 0.193 and 0.128, respectively.The gains reflect improved accuracy in both clinical direction and fine-grained change states.
  • Ablation: Removing the Temporal Change Agent sharply reduces LCC while leaving ReXrank performance largely unchanged.This attributes the LCC improvement to explicit temporal-change reasoning rather than general report quality.
  • Ablation: Replacing the base draft with the prior report reduces BLEU-1 from 0.466 to 0.381 but causes only a modest CE-F1 reduction.The base draft primarily supports report realization, whereas clinical decision agents determine core clinical content.

Conclusion

STRIVE is a multi-agent framework for longitudinal radiology report generation that assigns clinical reasoning to specialized agents and integrates progression-aware training with two-stage verification.

  • Framework: STRIVE assigns diagnosis, clinical attribute estimation, and temporal change reasoning to specialized agents rather than one model.This design separates the clinical reasoning tasks within the LRRG framework.
  • Temporal reasoning: Progression-Aware GRPO improves directional change modeling in STRIVE.The training method specifically targets temporal change reasoning.
  • Verification: The Consistency Gate and Validation Agent ensure consistency between structured clinical evidence and the generated report.These components provide two verification stages before and after report generation.
  • Evaluation: On Longitudinal-MIMIC, STRIVE improves report quality.The supplied conclusion passage reports this dataset-level outcome but does not provide the truncated numerical result.

Supplementary Material · S1 Dataset

Longitudinal-MIMIC pairs each current study with the patient’s most recent prior study and report, using patient-level partitioning to prevent overlap. Temporal comparisons appear in 55.0%–67.5% of reports across partitions, with 1.27–1.59 statements per report.

  • S1.1 Dataset Construction: Longitudinal-MIMIC is built from 26,625 MIMIC-CXR patients with two visits.
  • S1.1 Dataset Construction: Studies without a findings section are discarded before dataset partitioning.
  • S1.1 Dataset Construction: The dataset follows the official MIMIC-CXR patient-level split, ensuring no patient appears in more than one partition.
  • S1.1 Dataset Construction: Each example pairs a current study with the patient’s most recent prior study and its report.
  • S1.2 Temporal Change Statistics: Temporal comparisons occur in 55.0% of training reports, 56.2% of validation reports, and 67.5% of test reports.
  • S1.2 Temporal Change Statistics: Comparison statements have a density of 1.27 to 1.59 statements per report across the three partitions.

S2 Implementation Details

STRIVE uses frozen multimodal components with LoRA-adapted clinical decision agents and explicit training safeguards. Report generation combines structured clinical evidence with image details and retrieved style exemplars, followed by evidence-based validation.

  • Training Configuration: Diagnosis and Temporal Change Agents use Gemma-4-E4B-it, while the Attribute Agent uses MedGemma-1.5-4B-it with frozen backbones and completion-only bfloat16 LoRA.LoRA uses r=16, α=32, and dropout 0.05; checkpoints are selected on the patient-disjoint validation partition and training uses two NVIDIA RTX A6000 GPUs.
  • Chest X-ray Expert Pool: The Diagnosis Agent aggregates three classification experts and four generative experts, all frozen during training and evaluation.The classification pool comprises ConvNeXt, RAD-DINO, and CheXFound; the generative pool comprises MedGemma, PriorRG, CheXagent, and MAIRA-2.
  • Progression-Aware GRPO: The Temporal Change Agent receives SFT followed by progression-aware GRPO, sampling G=6 completions at temperature 1.0 without requiring a separately learned reward model.Group-relative advantages use rewards computed directly from sampled change states and report-derived targets.
  • Base Draft Generation: A frozen PriorRG model produces the base draft, while SCS findings retrieve the top K=3 training reports as style exemplars using IDF-weighted cosine similarity.Rare finding matches are favored and exemplars with extra findings are penalized through square-root IDF weighting.
  • Writer and Validation: The frozen Qwen3.6-27B Writer treats the SCS as authoritative clinical content, uses the base draft for image detail, and resolves conflicts in favor of the SCS.The Validation Agent applies frozen CheXbert labels, inserts omitted SCS-positive findings, and removes unsupported positives only when diagnosis probability is below 0.3.

S3 Evaluation Metrics

The evaluation combines surface-overlap, clinical-efficacy, semantic, and temporal metrics, while emphasizing that whole-report similarity can miss clinically decisive temporal errors. LCC directly evaluates stated temporal changes and distinguishes direction-preserving errors from reversals.

  • Surface Agreement: BLEU-n and ROUGE-L measure surface agreement with reference wording, using modified n-gram precision with brevity penalty and longest-common-subsequence F-measure, respectively.BLEU-n pools n-gram counts across 2,058 studies, whereas ROUGE-L is more recall-oriented because contiguity is not required.
  • Surface Agreement: Overlap metrics do not distinguish clinically decisive terms from stylistic words.This limitation motivates complementary clinical and temporal evaluation.
  • Clinical Efficacy: Clinical-efficacy Precision, Recall, and F1 use binarized CheXbert labels for 14 thoracic findings and are micro-averaged over finding–study pairs.Present is treated as positive, while absent, uncertain, and blank are treated as negative.
  • Semantic Evaluation: ReXrank extends evaluation beyond CheXbert’s 14-label constraint with study-level BLEU-2, BERTScore, and SembScore.These metrics capture bigram overlap, contextual-token similarity, and CheXbert-label embedding similarity.
  • Temporal Evaluation: LCC directly evaluates reference-stated temporal changes and distinguishes direction-preserving errors from reversals, unlike whole-report metrics that can obscure temporal mistakes.Removing the Temporal Change Agent substantially reduces LCC while having little effect on ReXrank.

S4 Longitudinal Change Concordance

Longitudinal Change Concordance (LCC) converts reports into structured finding-level change statements and compares generated reports with references using deterministic, reference-anchored matching. It measures temporal agreement while separating it from localization specificity.

  • Metric extraction: LCC uses the same instruction-tuned Gemma-4-31B, prompt template, and decoding configuration to extract change evidence from reference and generated reports.The model extracts finding-level evidence and labels but does not directly assign metric scores.
  • Metric extraction: Extraction returns compact statements containing an admissible finding, one of five change labels, evidence span, temporal cue, mention, laterality, and region.Reports without comparison cues receive an empty statement list through deterministic screening; the extractor does not infer change from current findings alone.
  • Matching and scoring: LCC performs reference-anchored, one-to-one matching by finding, while retaining laterality and region without requiring them as hard matching constraints.Compatible candidates are prioritized by fine-grained label agreement, isolating temporal agreement from localization specificity.
  • Matching and scoring: Macro-F1 is computed over one predicted label per reference statement at fine level across five labels and coarse level after collapsing new into increased and resolved into decreased.Support Devices and candidate-only statements are excluded from the primary reference-anchored score.

S5 In-Depth Analysis

The in-depth analysis shows that STRIVE substantially reduces omitted temporal changes, improves performance across fine-grained labels, and maintains temporal information through validation. Its gains are statistically distinguishable from baselines and are reflected in broader case-level coverage.

  • Temporal-change preservation: 0.46 omission rate: STRIVE reduces omitted reference changes by 0.25, or 35% relative to the best baseline.Baselines omit 0.71–0.87 of reference changes; removing the Temporal Change Agent raises omission to 0.83.
  • Per-label LCC decomposition: STRIVE achieves the highest per-label F1 for all five labels, with its largest gains on increased and resolved changes.Stable accounts for 2,219 of 3,275 reference changes, motivating macro averaging to avoid obscuring rarer directional states.
  • Statistical significance: STRIVE’s 95% confidence intervals lie entirely above those of all ten baselines at both LCC granularities.Using 2,000 paired nonparametric bootstrap resamples over 2,058 test study pairs, the strongest baseline’s LCC-C upper bound is 0.212 while STRIVE’s lower bound is higher.
  • Stage-wise pipeline analysis: 0.394/0.283 final LCC-C/F: Validation raises performance above the Writer’s 0.338/0.233 after structured states reach 0.354/0.250 through the Consistency Gate.The stage-wise trace identifies temporal information lost during verbalization and recoverable through validation.
  • Qualitative comparison: STRIVE covers every highlighted finding category in the additional case-level comparison, unlike prior methods that omit selected findings or focus primarily on stability and line placement.The categories are cardiomegaly, atelectasis, pneumothorax, pleural effusion, and support devices.

S6 Prompts

S6 specifies structured prompts for specialized agents that infer diagnosis, attributes, and temporal changes, then generate and validate reports. The schemas enforce explicit outputs, authoritative clinical state, and minimal-diff correction passes.

  • S6 Prompts: Diagnosis, Attribute, and Temporal Change Agents are fine-tuned for requested response formats, while the Writer and Validation Agent remain frozen.The prompts use inference-time slots and repeated per-finding blocks.
  • S6.1 Diagnosis Agent: The Diagnosis Agent integrates seven expert reads for each of 14 findings and assigns POS, UNC, or NEG in fixed-order JSON.Inputs combine three image-model probabilities and four CheXbert-style reads, while accounting for view-dependent visibility and disagreement.
  • S6.2 Attribute Agent: The Attribute Agent is queried once per positive finding to produce only severity, laterality, and anatomical-region modifiers, using open-vocabulary descriptors or “none”.It permits no output for absent findings and allows radiologist-natural modifiers beyond listed examples.
  • S6.3 Temporal Change Agent: The Temporal Change Agent combines direction, reports, time gap, diagnosis outputs, probabilities, interval deltas, and temporal dynamics to assign clinically notable change labels.It supports forward and time-reversed instances, outputs only JSON arrays, and assumes unlisted findings are unchanged.
  • S6.4 Writer: The Writer produces one continuous MIMIC-CXR-style paragraph whose present findings must exactly match the authoritative structured clinical state.Base drafts supply image details such as devices and measurements, while retrieved reports provide style only.
  • S6.4 Writer: Writer prompts require explicit positive findings, hedged uncertainty, and omission of absent findings while using natural radiology phrasing and continuity language.Severity terms are reserved for genuinely prominent new or worsening findings; stable findings use phrases such as “again seen” or “unchanged”.
  • S6.5 Validation Agent: The Validation Agent performs two passes: first reconciling finding presence, then inserting missing temporal states.Its minimal-diff editor applies only listed additions or removals and preserves all other report wording, order, and style.
  • S6.5 Validation Agent: Validation outputs only the corrected report as a single continuous paragraph after applying each specified correction.Corrections add state-supported findings missing from the report or remove asserted findings absent from the state under the stated probability condition.
Loading 2608.24237v1…