acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Towards Reliable, Auditable Clinical Evidence through Tool-Augmented Reasoning for Cine CMR Diagnosis

Abstract

Cine cardiac magnetic resonance (CMR) assesses cardiac structure and function. Despite advances in vision-language models (VLMs), automated diagnosis requires verifiable quantitative evidence and explicit checks of diagnostic sufficiency to avoid premature conclusions. We introduce TRACE (Towards Reliable, Auditable Clinical Evidence), a training-free agentic framework integrating adaptive tool use, clinical guideline retrieval, and independent evidence verification. Guided by clinical knowledge, the agent acquires CMR measurements, updates diagnostic hypotheses, and selects the next measurement. An independent verifier checks diagnostic requirements and returns evidence gaps for targeted acquisition and reassessment. Unresolved gaps lead to deferral for clinician review; accepted decisions produce structured reports linking findings, diagnostic rationale, clinical sources, and follow-up recommendations. On a development cohort of 50 balanced ACDC cases, TRACE achieves 87.3% mean accuracy across eight seeds with 96.3% diagnosis coverage, compared with 85.3% for ACDC-finetuned CineMA and 16%–28% for eight single-pass VLMs. Its report compiler achieves 100% specification conformance versus 46% for an LLM given the same patient evidence. These results demonstrate the value of evidence-guided agentic reasoning for auditable cine CMR diagnosis and reporting without additional model training.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
LLM Agents, Tool-Augmented Reasoning, Clinical Reasoning, Evidence Verification, Retrieval-Augmented Generation, Cardiac Magnetic Resonance

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LU7MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TRACE, a training-free agent framework for cine CMR diagnosis. It combines segmentation-derived measurements, clinical guidance, evidence verification, and structured reporting. The agent acquires evidence and revises diagnostic proposals after verification feedback. Accepted decisions are locked with their supporting evidence. The evaluation uses 50 balanced ACDC cases. TRACE achieves 87.3% accuracy and 96.3% diagnosis coverage across eight runs. Additional experiments examine acquisition policies, recovery after rejection, and report conformance.

Critical strengths

1.The framework achieves diagnostic performance comparable to a specialist classifier on the evaluated cohort without additional model training. These results support the feasibility of building effective diagnostic workflows from existing models, measurement tools, and clinical knowledge.

2.The audit follows initial proposals to final outcomes. Among 84 trajectories with rejected proposals, 76 end in correct diagnoses and eight end in deferral. The paper separately reports incorrect proposals accepted without rejection. This distinction makes the recovery results easier to interpret.

3.The paper includes single-pass reasoning, rule-based classification, and complete-evidence controls. These comparisons share the measurement backend. The statistical analysis accounts for repeated runs within patients and corrects for multiple comparisons. It helps the agent produce acceptable proposals with fewer revisions and model calls than static guidance. These results demonstrate a useful role for clinical grounding in guiding the reasoning process.

Critical weaknesses

1.The same 50 ACDC cases were used for development and evaluation. Their outcomes informed method refinement. No independent test cohort is reported. Repeated runs on these cases cannot establish generalization to new patients. The reported performance therefore remains limited to the development cohort.

2.TRACE achieves comparable diagnostic accuracy to the deterministic rule-based baseline (87.3% vs. 88.0%). Moreover, selective evidence acquisition performs worse than complete evidence exposure (86.7% vs. 92.0%). These results do not sufficiently demonstrate the additional benefits of the complex agentic workflow over a simpler rule-based system.

3.Section 4.2 reports that some MINF cases are misdiagnosed due to premature termination of evidence acquisition. Moreover, Section 4.3 reports that 12 incorrect initial diagnoses are directly accepted by the verifier. These results suggest that the current verification mechanism cannot consistently prevent unsupported or incorrect diagnoses, weakening the paper's central claim of improving diagnostic reliability.

Questions for authors

1.Which Evidence Contract thresholds or workflow rules were adjusted based on the 50 ACDC cases? Can the authors provide results on an independent test set with fixed rules to demonstrate generalization?

2.Can the authors provide a more comprehensive comparison of diagnostic accuracy, coverage, and computational cost to justify the additional complexity of the agentic workflow?

3.Why does the verifier accept some incorrect diagnoses without identifying missing evidence?

Reviewer_GRkXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TRACE combines segmentation-derived CMR measurements, clinical guidance, an adaptive language-model controller, and phenotype-specific evidence contracts. Verified decisions are locked before deterministic report compilation. On 50 ACDC development cases, it reports 87.25% accuracy and 96.25% diagnosis coverage, together with analyses of acquisition, recovery, and report conformance.

Critical strengths
  • Explicit separation of measurements, diagnostic proposals, verification, and report release gives a concrete and auditable implementation; unavailable measurements remain gaps rather than negative findings.
  • Shared-measurement controls, patient-clustered uncertainty, and transparent reporting of accepted errors and negative ablation results make the empirical analysis informative.
Critical weaknesses
  • All experiments use the same 50 cases that informed method refinement (Section 4; Appendix G). Repeated runs assess execution variability, not generalization to new patients. An untouched cohort with rules and prompts fixed beforehand is needed for the broader diagnostic claims.
  • The incremental value of adaptive reasoning remains limited: the deterministic reference obtains 88.0% accuracy, while complete evidence exposure reaches 92.0% versus 86.7% for selective acquisition and uses fewer model calls (Tables 2-3). Likewise, 100% report conformance verifies the implemented specification, not clinical appropriateness. These findings support an auditable implementation more strongly than a diagnostic advance.
Questions for authors
  1. Which contract thresholds, prompts, or control rules changed after examining outcomes on the 50-case cohort?
  2. Is there existing independent adjudication of the measurement proxies and contract acceptance criteria beyond agreement with ACDC phenotype labels?
Reviewer_unBROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TRACE uses a Qwen3-8B agent to request measurements from pretrained CineMA segmentations, consult fixed clinical cards, and propose one of five cine CMR phenotypes or a deferral without further training. A separate verifier applies evidence contracts before decisions and supporting evidence are locked for deterministic reporting and recommendation selection. On 50 balanced ACDC development cases, the authors report 87.25% mean accuracy and 96.25% coverage across eight runs. Recommendation evaluation measures specification compliance, without independent assessment of clinical appropriateness.

Critical strengths

The explicit separation of authority makes decisions traceable: the agent cannot write measurements or approve proposals, and reporting cannot alter accepted labels or evidence. Table 21 reports factual preservation across 400 executions; Table 23 separately checks software boundaries. The controls share segmentation masks, helping separate the measurement backend's contribution from the agentic workflow; this is more informative than the gap to image-only VLMs. Table 17 distinguishes process gains from diagnosis: dynamic guidance improves first-proposal acceptance and reduces calls without improving accuracy. Treating patients as independent units in the repeated-run analysis is appropriate.

Critical weaknesses

The central concern is that contract acceptance does not establish the diagnostic sufficiency motivating TRACE. Appendix C.2 allows DCM/RV acceptance without regional-motion assessment because unobserved competing predicates do not block release. A contract can therefore pass before evidence distinguishing MINF is collected.

In Table 3, complete exposure achieves 92.0% accuracy versus 86.7% with selective acquisition, using slightly fewer model calls. The authors attribute additional correct MINF classifications to obtaining motion before stopping. Separately, Appendix F.4 reports 12 wrong concrete diagnoses from five MINF patients, all accepted on their first proposal. The 84/84 recovery rate is conditional on trajectories containing a rejected proposal and does not cover these errors. The paper does not establish that every accepted error arose from missing motion, but the claim of preventing premature conclusions remains insufficiently supported.

Section 4 and Appendix G acknowledge that all 50 cases informed refinement; repeated runs do not test frozen rules on new patients. The deterministic rule reference achieves 88.0%; Table 16 estimates TRACE's difference as -0.8 percentage points, with a 95% interval of [-6.3, 5.0]. Equal accuracy can still offer auditability value, but the controller's additional diagnostic benefit remains uncertain without independent evaluation.

Questions for authors
  1. How often was motion unobserved in accepted DCM/RV predictions, and how do accuracy, coverage, and model calls change if available motion evidence is required before acceptance? Please provide the complete MINF support-index formula: Appendix B.3 gives ingredients, caps, and rounding, but omits the calculation needed to apply Table 11's 0.65 threshold.
  2. Can you evaluate cases excluded from refinement, with measurement rules, contracts, and thresholds frozen before evaluation, alongside the deterministic reference?

Comments on the paper

1 comment
Reviewer_LU7MOpenReview ↗

Response to author

Thank you for the detailed responses and new experiments. The ACDC-100 evaluation and hospital study provide stronger evidence for generalization. The motion-required experiment also helps explain the verifier’s limitations.

The results still do not show a clear benefit of the agent over a rule-based pipeline. On ACDC-100, TRACE achieves similar accuracy but requires about 13 model calls per case. Fewer measurements do not necessarily mean lower overall cost. The LGE results are encouraging. However, the hospital study lacks a rule-based baseline with the same tools, triage rules, verification contracts, and report compiler. This comparison is needed to assess the agent’s contribution. For this reason, I maintain my Weak Reject recommendation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.