acceptodds
Withdrawn from ICLR 2027

Cardiomni: An Evidence-Grounded Agentic Workflow for Multi-View Multi-Dimension Cardiac MRI Interpretation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cardiac magnetic resonance (CMR) interpretation spans a continuum from low-level visual perception to high-level clinical reasoning: a physician must inspect multi-view cine and LGE sequences, identify abnormal visual patterns, measure ventricular structure and function, and integrate these observations into a diagnosis. Existing CMR analysis systems are usually either task-specific predictors or end-to-end vision-language models that produce one-hop answers, making it difficult to accumulate explicit perceptual and quantitative evidence for downstream reasoning. Meanwhile, general and medical multimodal large language models have improved image-text understanding, but they are not designed to actively acquire CMR-specific evidence such as attribution maps, segmentation-derived volumes, ejection fraction, and indexed measurements. To bridge this gap, we propose Cardiomni, an evidence-grounded agentic workflow for multi-view, multi-dimensional CMR interpretation. Cardiomni coordinates two complementary evidence pathways through an iterative Planner–Executor–Verifier loop: a Visual Attention Highlighter provides abnormality probabilities with class-conditioned spatial evidence, while a Quantitative Measurement Estimator derives clinically interpretable measurements from cardiac masks and scan metadata. A persistent evidence ledger stores intermediate findings, enabling the planner to adaptively decide what evidence is still missing and allowing a frozen verifier and answer generator to produce answers grounded in the collected evidence. We train only the planner with reinforcement learning using a rule-based final-answer-correctness reward, while keeping the visual, quantitative, verification, and generation components fixed. We further introduce DeepCMRVQA, a benchmark for evaluating CMR perception-to-reasoning across 35 tasks covering major abnormality assessment and quantitative measurement interpretation. Experiments on binary abnormality VQA and measurement-range MCQ tasks show that Cardiomni outperforms general-purpose and medical vision-language baselines in balanced accuracy, highlighting a scalable path toward auditable CMR decision-support agents.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Agentic Workflow, Reinforcement Learning, Vision Language Models, Medical AI, Benchmark

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_9YGEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Motivated by the view that CMR interpretation is a multi-step reasoning process that gathers and integrates evidence across views and sequences, this paper proposes Cardiomni, an agentic workflow, as opposed to one-hop prediction models. Through a Planner–Executor–Verifier loop, Cardiomni iteratively calls the visual attention highlighter and the quantitative measurement estimator, records their outputs in an evidence ledger, and generates answers based on it. The authors also build DeepCMRVQA, a benchmark with 21 abnormality labels and 14 quantitative metrics, and report higher balanced accuracy than general and medical VLM baselines on binary VQA and quantitative MCQ.

Critical strengths

Strength: A CMR-specific agentic workflow that explicitly records its evidence

The paper presents an agentic workflow that addresses both disease assessment and quantitative measurement for multi-view, multi-sequence CMR of individual patients. Integrating abnormality classification and segmentation-based measurement as tools, and recording their outputs (classifier probabilities, attribution maps, measurements) in an evidence ledger so that the final answer is traceable, is a meaningful design.

Critical weaknesses

Weakness 1: The evidence selection mechanism lacks evaluation, implementation details, and explanation.

The paper emphasizes learning the evidence selection process (lines 69–70) and states that each question should be matched with evidence that can support it (lines 160–161). However, the reward for evidence selection is left as future work (lines 84–85, 303–306), and no quantitative evaluation of this process is provided. Since it is part of the core claims, deferring it to future work does not seem appropriate; observations and evaluation of the evidence-gathering process and its results for each task would be needed. It is also unclear how “predict relevant slices,” “retrieve additional slices,” and “refine segmentation” in Figure 3 actually work. Is “predict relevant slices” in the visual attention highlighter a separate task performed by the classification tool? Moreover, per Appendix B.2, ED/ES phases are first selected from segmentation results (the LV area curve of the mid slice), and then the full stack at those phases is segmented. This is not segmentation of selected slices, and the order appears reversed relative to Figure 3 (selection → segmentation). Do the slices chosen by “predict relevant slices” come from these ED/ES phases? Please also explain how views or sequences are selected.

Weakness 2: The experiments do not isolate the contribution of the agentic workflow.

That Cardiomni outperforms zero-shot models not specialized for cardiac MRI is expected, and the gain may come from the modality-specific tools themselves. To show the effect of the agentic workflow, comparisons with baselines that use the tool outputs directly are needed: for VQA, classifier probabilities with various thresholds (0.1, 0.2, …, 0.9), and for MCQ, measurements mapped directly to reference ranges. The fixed schedule in App. E.1 still goes through the LLM and cannot replace this. In addition, according to App. C, only the visual attention highlighter is active in VQA evaluation. The reported VQA results thus reflect repeated use of the same tool within the agent loop, not evidence pathway selection. The paper presents choosing one or both pathways within L_vol as a key decision (§3.1, lines 224–230), and Figures 1 and 3 show examples using both tools. However, this decision can only occur for L_vol (5 items), presumably MCQ items, and the reported experiments do not evaluate whether pathway selection is effective. Further experiments, such as reporting performance with only one tool enabled while keeping the Cardiomni loop (requiring separate training) versus both tools enabled, are needed to assess its contribution.

Questions for authors

Please see the Weaknesses section.

Reviewer_CM6TOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Cardiomni, a multimodal agentic framework for cardiac magnetic resonance (CMR) analysis. It uses a planner–executor–verifier loop and integrates two tool branches: classifier-based abnormality prediction with Grad-CAM attribution, and segmentation-based quantitative measurement using DICOM metadata. An evidence ledger records tool calls and results to connect final answers with intermediate evidence. The planner is trained with GRPO, while the other components remain frozen. The authors also introduce DeepCMRVQA, covering 21 binary abnormality-assessment tasks and 14 quantitative measurement-range selection tasks.

Critical strengths

Integrating CMR-specific tools has practical value. Combining visual attribution and segmentation-derived cardiac measurements supports multiple analysis tasks within one workflow and provides structured, inspectable evidence. DeepCMRVQA provides an initial benchmark for evaluating both abnormality assessment and quantitative interpretation. Its value to the community will depend on the availability of questions, labels, data splits, and a reproducible evaluation protocol. Cardiomni outperforms zero-shot general-purpose and medical multimodal models on the evaluated tasks. The fixed-schedule ablation also supports the usefulness of the adaptive loop as a whole.

Critical weaknesses

The evaluation mainly covers binary judgments about specified abnormalities and range selection for specified measurements. These tasks provide limited evidence for the broader claims about complex CMR interpretation and multistep clinical reasoning. It would be useful to test whether combining the two evidence branches also benefits patient-level disease classification, for example using suitable tasks from ACDC or M&Ms2. The introduction emphasizes learning to select between evidence branches. However, Appendix C states that VQA evaluation enables only the visual attribution branch, while MCQ evaluation uses the quantitative measurement branch. It remains unclear whether the planner can select appropriate evidence when both branches are available. Mixing VQA and MCQ questions while keeping both tools available would help assess this capability. Figure 3 also combines both branches, so its relationship to the main evaluation setting should be clarified.

Questions for authors

Although some data counts are reported, the numbers of independent patients, imaging sequences, and questions in each DeepCMRVQA split are not clearly summarized. Could the authors provide a table with these statistics for the VQA and MCQ subsets and clarify their relationship to the tool and planner training data? The question in Figure 1 explicitly names the target abnormality, which simplifies deciding what to investigate. Can the system also handle an open-ended request such as “Please analyze this CMR examination and report abnormalities and quantitative findings”? If this has not been evaluated, please clarify the scope of the claimed interpretation capabilities.

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

It proposes Cardiomni, an evidence-grounded agentic framework for cardiac MRI interpretation that integrates visual abnormality detection, quantitative measurement, and iterative reasoning through a Planner–Executor–Verifier workflow. Furthermore, it also introduces DeepCMRVQA, a benchmark covering 35 cardiac assessment tasks, and demonstrate improved performance over general-purpose and medical vision-language models using reinforcement learning to optimize evidence collection.

Critical strengths
  1. The integration of visual attribution, segmentation-based quantitative measurements, and a persistent evidence ledger provides a structured approach to cardiac MRI interpretation. This design improves the traceability of intermediate findings and aligns with clinical diagnostic workflows.

  2. Cardiomni achieves 73.0% balanced accuracy on binary VQA and 49.1% on quantitative MCQ tasks, outperforming the evaluated vision-language baselines. The improvements across diverse cardiac abnormalities and measurement metrics demonstrate the practical value of incorporating specialized imaging tools.

  3. The framework trains only the planner while keeping the perception, measurement, verification, and generation components frozen, offering a modular approach to agent optimization. The fixed-schedule ablation further demonstrates substantial improvements from the integrated adaptive workflow.

Critical weaknesses
  1. The evaluation enables only the visual pathway for VQA and only the quantitative pathway for MCQ, limiting evidence that the planner learns meaningful cross-pathway selection.

  2. The fixed-schedule ablation simultaneously removes learned planning, verification, iterative evidence collection, and reinforcement learning.

Questions for authors

Q1. Does the planner genuinely learn adaptive evidence routing? Since the current VQA and MCQ evaluations each enable only one evidence pathway, can the authors provide experiments requiring both visual and quantitative evidence?

Q2. Which components contribute most to the performance improvements? Could the authors provide separate ablations for GRPO training, evidence verification, iterative evidence collection, and learned routing?

Q3. How reliable are the quantitative measurements? Given the 49.1% balanced accuracy on quantitative MCQ tasks, can the authors provide more comprehensive measurement-error analysis against expert annotations or clinical reference values?

Q4. How is evidence sufficiency and citation correctness validated? Can the authors quantify how often the verifier incorrectly declares evidence sufficient or the generator produces conclusions unsupported by the evidence ledger?

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

Related papers

Discussion (0)

Sign in to comment.