Paper deep dive
CineMR: Tool-Integrated Vision-Language Reasoning for Quantitative Cardiac MRI Assessment
Kunyang Li, Hai Nguyen, Joshua Lowe, Chenguang Zhao, Peace C. Madueme, Mehdi Hedjazi Moghari, Mubarak Shah, Pegah Khosravi, Yuzhang Shang
Intelligence
Status: failed | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 0%
Last extracted: 10/6/2026, 10:16:09 AM
Unexpected non-whitespace character after JSON at position 192
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Cardiovascular magnetic resonance (CMR), including cine imaging, is a reference standard for the noninvasive assessment of cardiac morphology and ventricular function. Cine CMR interpretation integrates qualitative visual assessment with quantitative measurements of ventricular volumes, ejection fraction, myocardial mass, wall thickness, and regional wall motion. Current medical vision-language models (VLMs) cannot reliably derive quantitative measurements from multidimensional cine images without analysis tools. We present CineMR, a tool-augmented VLM that invokes cardiac image-analysis tools and integrates their outputs into interleaved reasoning for quantitative CMR assessment. We also construct a multi-cohort visual question answering benchmark covering quantitative metric extraction, multiclass diagnosis, and differential diagnosis, together with tools for segmentation, phase selection, volumetry, morphometry, and regional wall motion analysis. CineMR is trained with supervised fine-tuning (SFT) on tool-interaction traces followed by Group Relative Policy Optimization (GRPO) with conditional tool-use rewards. On the multi-cohort cine CMR benchmark, CineMR achieves 35.9% pass@1 and 58.9% pass@4, compared with 1.5% pass@1 for the Qwen3-VL-8B backbone and 0.0% and 7.0% pass@1 for LLaVA-Med v1.5 and MedGemma-4B, respectively. Correct tool invocation reaches 99.8% after GRPO, up from 78.9% after SFT. Live tool outputs improve ventricular measurement accuracy by 20.4--23.7% over direct model predictions, and removing all tools reduces pass@1 from 35.9% to 27.9%. These results highlight the importance of reliable tool use for quantitative cine CMR reasoning and support CineMR as a promising approach for assistive cardiac image assessment. Code, benchmark resources, and model weights are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2610.01166v2
- Canonical: https://arxiv.org/abs/2610.01166v2
Trouble viewing inline? Open PDF directly â
Full Text
75,015 characters extracted from source content.
Expand or collapse full text
CineMR: Tool-Integrated Vision-Language Reasoning for Quantitative Cardiac MRI Assessment Kunyang Li 1â Hai Nguyen 1,2â Joshua Lowe 1â Chenguang Zhao 3 Peace C. Madueme 3 Mehdi Hedjazi Moghari 4 Mubarak Shah 1§ Pegah Khosravi 1,2§ Yuzhang Shang 1§ 1 Institute for Artificial Intelligence, University of Central Florida 2 Department of Clinical Sciences, College of Medicine, University of Central Florida 3 Nemours Cardiac Center, Nemours Childrenâs Hospital, Florida 4 Childrenâs Heart Center, WVU Golisano Childrenâs, West Virginia Abstract Cardiovascular magnetic resonance (CMR), including cine imaging, is a reference standard for the noninvasive assessment of cardiac morphology and ventricular function. Cine CMR interpretation integrates qualitative visual assessment with quantitative measurements of ventricular volumes, ejection fraction, myocardial mass, wall thickness, and regional wall motion. Current medical vision-language models (VLMs) cannot reliably derive quantitative measurements from multidimen- sional cine images without analysis tools. We present CineMR, a tool-augmented VLM that invokes cardiac image-analysis tools and integrates their outputs into interleaved reasoning for quantitative CMR assessment. We also construct a multi-cohort visual question answering benchmark covering quantitative metric extraction, multiclass diagnosis, and differential diagnosis, together with tools for segmentation, phase selection, volumetry, morphometry, and regional wall motion analysis. CineMR is trained with supervised fine-tuning (SFT) on tool- interaction traces followed by Group Relative Policy Optimization (GRPO) with conditional tool-use rewards. On the multi-cohort cine CMR benchmark, CineMR achieves 35.9% pass@1 and 58.9% pass@4, compared with 1.5% pass@1 for the Qwen3-VL-8B backbone and 0.0% and 7.0% pass@1 for LLaVA-Med v1.5 and MedGemma-4B, respectively. Correct tool invocation reaches 99.8% after GRPO, up from 78.9% after SFT. Live tool outputs improve ventricular measurement accu- racy by 20.4â23.7% over direct model predictions, and removing all tools reduces pass@1 from 35.9% to 27.9%. These results highlight the importance of reliable tool use for quantitative cine CMR reasoning and support CineMR as a promising approach for assistive cardiac image assessment. Code, benchmark resources, and model weights are available at https://github.com/AI-MIND-Lab/CineMR. 1 Introduction Cardiovascular magnetic resonance (CMR) is a reference standard for the noninvasive quantification of ventricular volumes, myocardial mass, and systolic function [Petersen et al., 2016, Schulz-Menger et al., 2020]. Clinical CMR interpretation integrates visual assessment with quantitative measure- ments of ventricular volumes, ejection fraction, myocardial mass, wall thickness, and regional wall motion, together with clinical history and complementary imaging findings [Arbelo et al., 2023, Schulz-Menger et al., 2020]. Quantitative criteria can support phenotype classification, but definitive diagnosis may also require clinical, electrocardiographic, genetic, coronary, and tissue- characterization information. The combination of multidimensional image analysis, quantitative measurement, and criterion-based interpretation makes CMR a compelling testbed for tool-augmented Preprint. â Equal contribution. § Corresponding authors. arXiv:2610.01166v2 [cs.CV] 5 Oct 2026 <think>There are no obvious wall motion abnormalities in this single slice. However, I need to consider the possibility of regional wall motion abnormalities in other slices. I will also need to consider the patient's clinical history and other imaging findings to arrive at a definitive diagnosis.</think> <think>The patient's cardiac MRI shows normal left and right ventricular volumes and wall thickness. There are no obvious wall motion abnormalities in this single slice. However, I need to consider the possibility of regional wall motion abnormalities in other slices and the patient's clinical history and other imaging findings to arrive at a definitive diagnosis. I will need to review the entire cardiac MRI to assess for any regional wall motion abnormalities and to evaluate the patient's clinical history and other imaging findings.</think> [... identical <think> block repeated 15 more times ...] <think>The patient's REPETITION GENERATION Question: Analyze this cardiac cine MRI and provide a comprehensive diagnostic assessment. Systematically evaluate the left and right ventricular morphology and function, identify any regional wall motion abnormalities, and arrive at the most likely diagnosis. Use the hypothesis verification approach: state your initial phenotypic hypothesis, then verify with quantitative measurements. Reasoning (text only): To analyze this cardiac cine MRI, I would first observe the left and right ventricular morphology and function. This includes assessing the ventricular chamber sizes, ejection fraction (EF), and any regional wall motion abnormalities. Based on these observations, I would formulate a preliminary phenotypic hypothesis ... Next, I would verify the hypothesis using quantitative measurements obtained from the MRI images. For example, I might use tools to segment the cardiac structures, ..., and analyze the trabeculation patterns. By comparing these measurements ... NO DEFINITIVE ASSESSMENT PRODUCED LLaVA-Med v1.5 (Mistral-7B)MedGemma (Google, 4B) <think>The image shows a single axial slice of the heart. I need to assess the left and right ventricular morphology and function. I will start by segmenting the cardiac structures and then measuring the volumes and wall thickness. Based on these measurements, I can formulate a hypothesis about the patient's cardiac condition and then verify it with additional imaging or clinical information.</think> <tool_call>"name": "segment_cardiac", "arguments": "image": "image.png"</tool_call> <tool_result>"left_ventricle": "volume": 60, "wall_thickness": 12, "right_ventricle": "volume": 30, "wall_thickness": 8, "atria": "left_atrium": "volume": 20, "right_atrium": "volume": 10</tool_result> <think>The left ventricle volume is 60 ml, and the right ventricle volume is 30 ml. The left ventricular wall thickness is 12 m, ..., and the right atrium volume is 10 ml. These measurements are within normal limits. I will now analyze the wall motion abnormalities.</think> UNSUPPORTED TOOL OUTPUT CineMR (Ours, Based on Qwen3-VL-8B) <think> Hypothesis: The imaging pattern suggests a dilated, globally hypokinetic phenotype. I need to quantify these observations with tools.</think> <think> I need to segment the cardiac structures to identify the left ventricle, right ventricle, and myocardium.</think> <tool_call> "name": "segment_cardiac", "arguments": </tool_call> <tool_result> "status": "success", "segmentation": " <seg_mask>"</tool_result> <think> Segmentation complete. I can now compute quantitative measurements.</think>... <think> Verification 1: LVEF = 29.8572637810542% â severely reduced (normal: 52â72%). ... Conclusion: DCM-compatible phenotype under the benchmark criteria.</think> Figure 1: Comparison of model outputs for the same cine CMR case, where the benchmark reference category is dilated cardiomyopathy (DCM). In this illustrative example, LLaVA-Med fails to provide a definitive assessment, whereas MedGemma produces unsupported tool-call content and repetitive text. In contrast, CineMR invokes executable tools to obtain quantitative measurements and evaluates them according to criteria encoded in the study-specific Rule-Based Cardiac Phenotype Classifier, thereby producing a traceable output grounded in explicit measurements and decision rules. multimodal reasoning. This is particularly important because errors in upstream measurements can propagate to downstream classifications. Recent medical VLMs have expanded multimodal question answering and generated-rationale capabilities, with reinforcement learning to optimize structured response generation [Lai et al., 2025, Pan et al., 2025, Su et al., 2025, Li et al., 2023, Sun et al., 2025]. However, general-purpose medical VLMs are not designed to perform the multidimensional image analysis required to derive quantitative CMR measurements reliably. As illustrated in Fig. 1, the evaluated baseline models may fail to produce a definitive assessment or generate repetitive text, unsupported tool outputs, or ungrounded quantitative values when prompted to interpret cine CMR. This limitation reflects the explicit multidimensional computations required for quantitative CMR assessment. For example, assessment of a DCM-compatible phenotype requires delineation of the left ventricle across multiple slices and cardiac phases, voxel-based calculation of ventricular volumes and ejection fraction, and interpretation using reference ranges and study criteria [Petersen et al., 2017]. Current VLMs cannot be assumed to reproduce this multistep quantitative pipeline reliably from image inputs alone. Tool-augmented medical agents [Li et al., 2024, Fathi et al., 2025, Wang et al., 2026, Jiang et al., 2025] offer a promising direction, although many existing systems emphasize perceptual operations, such as segmentation visualization and region enlargement, rather than the explicit computation of clinically interpretable imaging measurements. Quantitative descriptors such as the septal-to- lateral wall-thickness ratio and mass-to-volume ratio may contribute to differentiating hypertrophic cardiomyopathy (HCM) from hypertensive heart disease (HHD), but they must be interpreted with additional morphological and clinical information; assessment of a DCM-compatible phenotype incorporates ventricular dilation and systolic dysfunction relative to appropriate sex-specific reference ranges [Petersen et al., 2017]. These measurements cannot be quantified reliably through visual inspection alone. These tasks require quantitative measurement in addition to visual interpretation, motivating tools that return explicit quantitative outputs. CineMR addresses this need by equipping a VLM with cardiac image-analysis tools and training it to invoke and integrate their outputs within an interleaved reasoning process. Rule-Based Cardiac Phenotype Classifier-structured dataset. To make the supervision clinically traceable and systematically cover criterion-based reasoning, we organize the benchmark around a Rule-Based Cardiac Phenotype Classifier rather than relying only on direct image-to-label associa- tions. The Rule-Based Cardiac Phenotype Classifier provides explicit diagnostic pathways linking quantitative measurements to reference answers and expected tool use. We derive the training data from a study-specific Rule-Based Cardiac Phenotype Classifier informed by published cardiomyopa- thy criteria, CMR reference studies, and explicitly documented project-specific thresholds [Arbelo et al., 2023, Petersen et al., 2017]. The Rule-Based Cardiac Phenotype Classifier covers ten harmo- nized outcome categories. Each question maps to a Rule-Based Cardiac Phenotype Classifier node annotated with its task-complexity level (L1âL6) and the tools expected to generate the reference an- swer. This structure encourages the model to follow criterion-based pathways and provides traceable supervision for both answer generation and tool use. 2 Clinical Evaluation History, syptoms, and risk factors CMR Acquisition Multi-parametric, standard protocol Physician Analysis Visual assessment, automated calculation and physician verification Clinical Decision Structured report and management plan Real-world Clinical Workflow â Input Short-axis Cine CMR Sequence time Question: Assess cardiac morphology, function and identify the most likely benchmark category ⥠Vision-Language Model Qwen3-VL-8B-Instruct Multimodel Reasoning and Generation ⣠Rule-Based Cardiac Phenotype Classifier Criterion-Based Classification Using Encoded Rules LVEDVi: 135.4 mL/m 2 LVEF: 23.7% 57.4% Quantitative Cardiac Metircs global hyperkineticMotion: RVEF: MVR:0.533 g/mL ...... 10-category cardiac phrnotype classification †Output Predicted Cardiac Phenotype: DCM-Compatible Phenotype Quantitative Evidence LVEF: 23.7% (reduced) LVEDVi: 135.4 ml/m 2 (above the applicable reference range) Global hypokinesis No significant RV involvement Compatible with a DCM phenotype; etiology is not established from cine imaging alone. Structured Model Output Morphology: LV, RV and myocardial geometry Function: Temporal ventricular dynamics Initial Imaging Hypothesis: Quantitative Functional Assessment âą Tool-Augmented Analysis 3.1 Cardiac Segmentation (Tool 1) ED frameES frame RV MYO LV 3.2 Ventricular Volumetry (Tool 2) ED 295 mL ES 225 mL Output: LV/RV/Myocardial masks Output: EDV, ESV, EF 3.3 Regional Wall Motion (Tool 3) Output: Regional Wall Motion Normal Hypokinetic Akinetic CineMR Framework One-to-one Correspondence Figure 2: Overview of the CineMR framework, designed to mirror the real-world cardiologist workflow. Just as a clinician first inspects the cine images to form a visual impression, CineMR generates an initial phenotypic hypothesis via chain-of-thought reasoning. The clinician then orders quantitative analyses (segmentation, volumetry, wall thickness measurement); analogously, CineMR invokes a cardiology-specific quantitative tool suite to obtain precise metrics (EDV, ESV, EF, mass, MVR, wall thickness, sphericity index, regional wall motion). Finally, where the clinician interprets these measurements against ESC/AHA guideline thresholds to reach a diagnosis, CineMR routes them through a guideline-aligned Rule-Based Cardiac Phenotype Classifier with threshold-gated branching, producing an auditable reasoning chain that parallels the clinical diagnostic process from visual assessment to quantitative verification to evidence-based conclusion. Cardiac image-analysis tool suite. We develop a cardiology-specific tool suite for quantitative cine CMR analysis. LV, RV, and myocardial segmentation is performed using a task-specific model fine- tuned from CineMA [Fu et al., 2025]. Voxel-based volumetry computes biventricular end-diastolic volume (EDV), end-systolic volume (ESV), and ejection fraction (EF). Morphometric analysis further derives myocardial mass, mass-to-volume ratio (MVR), maximal wall thickness, septal-to-lateral wall-thickness ratio, and sphericity index. Regional wall motion is assessed at the segment level using an AHA segment-based representation. These tools provide the VLM with segmentation-derived quantitative measurements. The measurements are used at the principal quantitative branches of the study-specific Rule-Based Cardiac Phenotype Classifier. This design allows the model to condition its responses on explicit quantitative evidence rather than on image features alone. Two-stage training with conditional tool rewards. A conventional approach to tool-use training is supervised fine-tuning (SFT) on reference tool-interaction traces. SFT teaches the model the tool-call syntax, the interleaved reasoning format, and the association between different question types and their expected tools. However, SFT mainly imitates predefined trajectories and does not directly optimize whether the model selects tools appropriately during its own generation. We therefore further apply Group Relative Policy Optimization (GRPO) [Shao et al., 2024] to refine the modelâs tool-use policy. The GRPO stage uses a conditional tool-use reward, assigning credit for tool invocation only 3 Phase 1: Initial Imaging Hypothesis <think> Hypothesis: The imaging pattern suggests a dilated, globally hypokinetic phenotype. I need to quantify these observations with tools.</think> Phase 2: Tool-Integrated Quantitative Analysis <tool_call> "name": "segment_cardiac", "arguments": </tool_call> <tool_result> "status": "success", "segmentation": " <seg_mask>"</tool_result> Phase 3: Criterion-Based Classification <think> Verification 1: LVEF = 29.8572637810542% â severely reduced (normal: 52â72%). Verification 2: LVEDV = 201 mL â dilated (>150 mL absolute threshold). Verification 3: MVR = 0.424 g/mL (ACDC-calibrated). Low (<0.49) â eccentric remodeling. Verification 4: Max EDWT = 10.2 m â no hypertrophy (<13 m). Verification 5: No focal regional WMA â dif fuse global hypokinesis (14/16 segments affected). Against ischemic etiology (IHD typically shows focal/territorial WMA). Verification 6: SI = 0.722 â pathological sphericization (>0.57). [Maceira 2006: normal males 0.35±0.06.] Verification 7: RVEDV/LVEDV = 0.72 â LV-dominant dilation (not balanced biventricular as in AHS). Verification 8: Maximal apical FD = 0.812 â LVNC not supported (<1.3). Decision path: FD_check:FD=0.812<1.3(not_LVNC) â LVEF=29.9%(severely_reduced,<40%) â LVEDV=201mL(dilated,>150mL) â reduced_EF+dilated_LV â EDWT=10.2m<13m(no_hypertrophy) â SI=0.722>0.57(pathological_sphericization) â diffuse_hypokinesis(14/16_segments) â MVR=0.424<0.49(eccentric_remodeling) Conclusion: DCM (High confidence).</think> Verbatim Model Output The diagnosis is Dilated cardiomyopathy (DCM). LVEF 29.8572637810542% with LVEDV 201 mL (dilated) and MVR 0.424 g/mL (ACDC-calibrated). Diffuse global hypokinesis (14/16 segments affected). Pathological sphericization (SI=0.722). No hypertrophy. Confidence: High. DCM EDESEDES segmentation output <think> Segmentation complete. I can now compute quantitative measurements.</think> <think> I need to segment the cardiac structures to identify the left ventricle, right ventricle, and myocardium.</think> <tool_call> "name": "measure_volume", "arguments": </tool_call> <tool_result> "lv_edv_ml": 201.38291534087625, "lv_esv_ml": 141.25837209463518, "lv_ef_pct": 29.857263781054238, "rv_edv_ml": 145.73829104625381, "rv_esv_ml": 65.41927385016429, "rv_ef_pct": 55.112048372591846, "lvedvi_ml_m2": null, "rvedvi_ml_m2": null, "rvedv_lvedv_ratio": 0.7237461825394107 </tool_result> <think> Volume measurements obtained. Let me compute additional cardiac metrics.</think> <tool_call> "name": "compute_cardiac_metrics", "arguments": </tool_call> <tool_result> "lv_mass_g": 85.29174036281547, "mvr_acdc_g_ml": 0.4235074918362547, "max_edwt_m": 10.2, "sl_ratio": 1.0174628391045237, "septal_convexity_m": 0.0, "sphericity_index": 0.7218463051827394, "max_apical_fd": 0.8124736509281475, "has_focal_wma": false, "wma_territory": "none", "diffuse_hypokinesis": true, "akinetic_segments": [1, 6], "hypokinetic_segments": [2, 3, 4, 5, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16] </tool_result> Question: Analyze this cardiac cine MRI and provide a comprehensive diagnostic assessment. Systematically evaluate the left and right ventricular morphology and function, identify any regional wall motion abnormalities, and arrive at the most likely diagnosis. Use the hypothesis verification approach: state your initial phenotypic hypothesis, then verify with quantitative measurements. CineMR: Tool-Integrated Reasoning for Quantitative CMR Assessment Verbatim Benchmark Prompt Figure 3: Representative verbatim CineMR rollout for a cine CMR case assigned to the DCM category. The figure shows the original benchmark prompt, model-generated tool calls, executable tool outputs, intermediate generated rationale, and final predicted category. CineMR invokes cardiac segmentation, ventricular volumetry, and morphometric analysis tools before applying the published and study-specific criteria encoded in the Rule-Based Cardiac Phenotype Classifier. Tool results are displayed at their native quantitative precision. The intermediate rationale and final response are reproduced verbatim and should not be interpreted as independently validated clinical conclusions. when the final answer is correct. This encourages the model to learn when tool use is beneficial while reducing unnecessary or indiscriminate tool calls. We evaluate CineMR on a multi-cohort benchmark constructed from ACDC [Bernard et al., 2018], M&Ms, and M&Ms-2 [Campello et al., 2021, MartĂn-Isla et al., 2023]. The benchmark spans ten harmonized outcome categories and six levels of task complexity (L1âL6). CineMR achieves 35.9% pass@1 and 58.9% pass@4 overall, compared with 1.5% pass@1 (4.5% pass@4) for the unfine- tuned Qwen3-VL-8B backbone, 7.0% pass@1 (18.0% pass@4) for MedGemma-4B-it, and 0.0% pass@1 (0.0% pass@4) for LLaVA-Med v1.5. On L3âL4 tasks requiring criterion-based quantitative assessment, CineMR achieves 58.7% pass@1 and 85.5% pass@4, whereas end-to-end multicategory diagnosis across all ten benchmark categories (L5) remains substantially harder, at 13.3% pass@1, indicating that CineMR is not yet a reliable standalone diagnostic system. GRPO also raises correct tool selection to 99.8% of gold tool-requiring items, up from 78.9% after supervised fine-tuning, and live tool outputs improve ventricular measurement accuracy by 20.4â23.7% over direct model predictions, highlighting the importance of reliable tool use for quantitative cine CMR reasoning and supporting tool-augmented VLMs as a promising approach for assistive cardiac image assessment. 2 Related Work Tool-Augmented Medical Agents.Integrating multimodal large language models (MLLMs) with external tools can extend their capabilities beyond image and text generation. MMedAgent [Li et al., 2024], VILA-M3 [Nath et al., 2025], and AURA [Fathi et al., 2025] incorporate task-specific tools for operations such as segmentation, classification, and image annotation. MedAgent-Pro [Wang et al., 2026] and AgentMD [Jin et al., 2024] organize specialized models or tools within agent-based diagnostic or clinical risk-assessment workflows. Ophiuchus [Jiang et al., 2025] is particularly relevant because it uses a three-stage training strategy comprising supervised fine-tuning, self- reflection fine-tuning, and agentic Group Relative Policy Optimization [Shao et al., 2024] to train a 4 VLM to select and invoke image-analysis tools within an interleaved generation process. Relative to the requirements of quantitative CMR assessment, many of these approaches emphasize perceptual tools, such as segmentation visualization and region enlargement, rather than tools that return explicit imaging measurements. In addition, several systems rely primarily on prompting or supervised fine-tuning for tool selection. CineMR extends this paradigm by introducing quantitative tools that derive explicit quantitative measurements from segmentation outputs and by organizing training data around a study-specific Rule-Based Cardiac Phenotype Classifier designed to cover the encoded benchmark criteria systematically. AI for Cardiac MRI Analysis.Deep learning methods have shown strong performance on several cine CMR analysis tasks. The ACDC challenge [Bernard et al., 2018], M&Ms challenge, and M&Ms-2 challenge [Campello et al., 2021, MartĂn-Isla et al., 2023] provide widely used benchmarks for cardiac segmentation. Representative segmentation approaches include nnU-Net [Isensee et al., 2021] and task-specific models fine-tuned from the CineMA foundation model [Fu et al., 2025]. Prior work has also developed automated pipelines for quality-controlled segmentation and quantitative characterization of cardiac function [Ruijsink et al., 2020], as well as radiomics models for char- acterizing cardiovascular risk-related imaging phenotypes [Cetin et al., 2020]. To our knowledge, prior work has not combined executable quantitative cine CMR analysis tools with an interleaved VLM generation process for criterion-based cardiac phenotype assessment. CineMR differs from these approaches by training a VLM to generate an initial imaging hypothesis, invoke executable tools for quantitative analysis, and integrate the returned measurements using criteria encoded in a study-specific Rule-Based Cardiac Phenotype Classifier. 3 Method CineMR trains a vision-language model to perform criterion-based cardiac phenotype assessment by strategically invoking quantitative tools within an interleaved generation process. An overview of the proposed framework is shown in Fig. 2. We first formulate the diagnostic reasoning problem (§3.1), and then introduce the Rule-Based Cardiac Phenotype Classifier-structured dataset (§3.2), the quantitative tool suite (§3.3), and the two-stage training pipeline (§4). 3.1 Problem Formulation LetIdenote a cine CMR sequence,qa benchmark question, andT = t 1 ,...,t K the available tool set. We model the task as an agentic generation process: at each turni, the model produces reasoning textr i or emits a tool invocationÏ i = (t k ,a i ); the runtime executest k on the imaging input I using argumentsa i and returnso i = t k (I,a i )which is inserted into the generation context. The full responsey = (r 1 ,Ï 1 ,o 1 ,r 2 ,...,r n )ends with a predicted answer Ë d â D, whereDmay represent a numerical value, binary finding, or benchmark category, depending on the question type. As shown in Fig. 3, this proceeds in three phases: (1) a<think>block forming an initial imaging hypothesis, (2) one or more<tool_call>/<tool_result>turns obtaining quantitative measure- ments, and (3) a narrative synthesis that evaluates the returned measurements using the published and study-specific criteria encoded in the Rule-Based Cardiac Phenotype Classifier before producing the final predicted category Ë d. 3.2 Rule-Based Cardiac Phenotype Classifier-Structured Dataset Published cardiomyopathy guidelines and CMR reference studies contain quantitative criteria that can inform phenotype assessment. We encode a subset of these criteria, together with explicitly documented study-specific thresholds, into benchmark decision rules, including criteria related to LVEF, wall thickness, ventricular volumes, and regional wall motion. Several CMR findings relevant to cardiac phenotype assessment can be represented using quantitative criteria. However, definitive clinical diagnosis also incorporates history, symptoms, electrocardiography, genetics, loading conditions, coronary assessment, and tissue-characterization imaging. This design provides three advantages: (i) criterion-traceable supervision based on deterministic answer-generation rules rather than unconstrained LLM-generated labels, (i) node-level annotations specifying the tools expected by the reference policy, and (i) six task-complexity levels (L1âL6). 5 CMR Analysis: LVEF, Volumes, Mass, EDWT, Wall Motion Apical FD â„ 1.30? LVNC Probable Yes ARVC Screen Modified Marcus 2010 TFC RVEF â€40% or RVEDVi â„110/100 + RV regional WMA ARVC No Criteria met + disproportionate RVEDV/LVEDV > 1.2 LVEF Not met LVEF â„ 52% (Preserved) 40% †LVEF < 52% (Mildly Reduced) LVEF < 40% (Reduced) â„ 52% 40-52% < 40% Max EDWT â„ 15m? S:L ratio > 1.3? â„ 15m (Hypertrophy) MVR > 1.14 (ACDC-cal)? Yes (Asymmetric) HCM High HCM Probable > 1.14†1.14 HHD No (Symmetric) LVEDVi Elevated? < 15m NOR High Normal AHS Evaluation SI < 0.57, MVR 0.54-0.78 RVEDV/LVEDV ~1.0 Elevated AHS INDETERMINATE Criteria met Not met LVEDVi > 100 mL/mÂČ? INDETERMINATE Reduced EF, non-dilated Normal (non-dilated) Focal WMA? Dilated EDWT †5.5m in WMA? Yes (Focal) MINF Probable IHD Probable †5.5m (Scar) > 5.5m (Viable) MVR Evaluation SI, MVR, RVEDV/LVEDV No (Diffuse / None) MVR < 0.49 (ACDC-cal)? DCM Yes + SI > 0.57 MVR 0.54-0.78 + balanced RV/LV? No (â„ 0.49) AHS INDETERMINATE Yes No Rule-Based Cardiac Phenotype Classifier Figure 4: The tools produce the quantitative measurements required by the study-specific Rule-Based Cardiac Phenotype Classifier. Final classification includes the following categories: hypertrophic cardiomyopathy (HCM), athleteâs heart or physiological athletic remodeling, arrhythmogenic right ventricular cardiomyopathy (ARVC), left ventricular non-compaction (LVNC), ischemic heart disease (IHD), myocardial infarction (MINF), dilated cardiomyopathy (DCM), hypertensive heart disease (HHD), normal (NOR), and indeterminate. In clinical practice, CMR interpretation integrates visual assessment with quantitative measurements and other available clinical and imaging information rather than exhaustively traversing a fixed Rule-Based Cardiac Phenotype Classifier [Schulz-Menger et al., 2020]. The decision rules represent the verification pathways encoded for this benchmark, while the model learns to select tools and integrate their outputs using the corresponding published and study-specific criteria. For each patient, routing the extracted metrics through the Rule-Based Cardiac Phenotype Classifier yields a candidate diagnosis, which is reconciled against the cohortâs independently confirmed clinical label: whenever the two disagree, the confirmed cohort label takes precedence and becomes the boxed reference answer. This reconciliation step overrode the classifierâs candidate diagnosis in only 2 of 3,320 patients, indicating that the encoded rules agree with the cohortâs confirmed diagnoses in the overwhelming majority of cases. The Rule-Based Cardiac Phenotype Classifier in Fig. 4 assigns cases to ten harmonized benchmark categories: NOR, DCM, HCM, HHD, MINF, IHD, ARVC, athleteâs heart, LVNC, and INDETERMI- NATE. Each branching point uses either a published quantitative criterion or an explicitly documented project-specific threshold or calibration. LVEF classification uses sex-specific reference ranges re- ported by [Petersen et al., 2017]. The HCM-related benchmark rules are informed by the 2023 ESC cardiomyopathy guideline [Arbelo et al., 2023]. In adults, otherwise unexplained maximal LV wall thickness ofâ„15 m supports HCM, whereas a threshold ofâ„13 m may support HCM in first-degree relatives of patients with unequivocal HCM or in the presence of other compelling clinical or genetic evidence. The ARVC-related benchmark rules use the structural and functional CMR components of the 2010 Revised Task Force Criteria [Marcus et al., 2010]. For ACDC, MVR thresholds are adjusted using a project-specific factor of 0.913 to account for differences in the treatment of papillary muscles. This adjustment is informed by studies demonstrating the effects of papillary-muscle inclusion on LV mass and volume measurements [Vogel-Claussen et al., 2006, Gommans et al., 2016]. No corresponding adjustment is applied to M&Ms or M&Ms-2. Task-complexity levels. We organize questions into six cognitive layers of increasing reasoning complexity: 6 Table 1: Cardiology-specific tool suite. Each tool maps to Rule-Based Cardiac Phenotype Classifier nodes requiring quantitative thresholds inaccessible to visual inspection. ToolInputâ OutputUnlocked MetricsRole segment_cardiacI âsLV/RV/MYO masksprerequisite measure_volumesâvEDV, ESV, EF, EDVicomputational compute_cardiac_metrics(s,v)âmmass, MVR, EDWT, S:L, SIcomputational wall_motion_analysissâwsegment-level WMAcomputational select_cardiac_phasessâ(t ED , t ES )temporal groundingauxiliary zoom_anatomy(t, target)â ROIlocal visual detailauxiliary âąL1, numeric extraction: direct retrieval of quantitative metrics, e.g., âWhat is the LVEF?â Tool use is required. âąL2, binary finding assessment: evaluation of a visual or anatomical finding, e.g., âIs the LV dilated?â Tool use is optional under the benchmark reference policy, except for templates that explicitly require quantitative measurements. âąL3, criterion evaluation: comparison with a published or study-specific threshold, e.g., âDoes EDWT satisfy the HCM criterion ofâ„ 15 m?â Tool use is required. âąL4, pairwise criterion-based classification: distinction between two benchmark categories, e.g., HCM versus HHD. Multiple tools are required under the reference policy. âąL5, multicategory assessment: end-to-end prediction across all ten benchmark categories using a hypothesis-and-verification format. âąL6, targeted differential assessment: distinction between clinically confusable benchmark categories, e.g., DCM versus MINF. The resulting 19 templates comprise 6, 4, 4, 2, 1, and 2 questions for L1, L2, L3, L4, L5, and L6, respectively.(Appendix B lists them all). These levels progress from quantitative measurement and criterion evaluation to multicategory and targeted differential assessment, with L3âL6 requiring increasingly selective and compositional tool use under the benchmark reference policy. Metric provenance. We audited all 66 classifier thresholds for clinical traceability: 43 were grounded in guidelines or peer-reviewed studies, while 23 were project-specific thresholds with documented derivations and published precedents. 3.3 Cardiology-specific Quantitative Tool Suite Clinical cardiac MRI diagnosis is fundamentally a measurement-driven process: ESC/AHA guidelines define diagnostic criteria through quantitative thresholds, such as LVEF< 40%for reduced systolic function and EDWTâ„ 15mmfor HCM [Arbelo et al., 2023, Ommen et al., 2020]. Obtaining these measurements from raw imaging data requires multiâstep computation, including anatomical delineation, voxel-wise integration, and derived ratio calculation, which current VLMs cannot reliably perform internally. Many prior tool-augmented medical agents [Li et al., 2024, Jiang et al., 2025] emphasize perceptual tools, such as segmentation visualization and region enlargement, rather than tools that return explicit quantitative imaging measurements. We therefore design a quantitative tool suite withK = 6tools, organized as a dependency chain in which each toolâs output feeds subsequent computations. These tools produce the quantitative measurements required by the study-specific Rule-Based Cardiac Phenotype Classifier. Segmentation tool.segment_cardiac(I) âsuses a task-specific segmentation model fine- tuned from CineMA [Fu et al., 2025], a multiview convolutional-transformer masked-autoencoder foundation model pretrained on approximately 15 million cine CMR images from 74,916 UK Biobank studies, to produce frame-wise masks for the left ventricle (LV), right ventricle (RV), and myocardium (MYO). This tool is a prerequisite for downstream quantitative analyses, whose measurements are derived using standardized quantitative post-processing rather than unstructured visual estimation. Volumetric tools.measure_volume(s) âvcomputes computes biventricular end-diastolic and end-systolic volumes, ejection fractions, and, when body surface area is available, indexed volumes using voxel-based volumetry across the short-axis stack. When body surface area is unavailable, indexed volumes are not reported. LVEF and indexed ventricular volumes are major quantitative features used by the study-specific Rule-Based Cardiac Phenotype Classifier [Arbelo et al., 2023]. 7 LVEF determines the initial branch preserved, mildly reduced, or reduced while indexed volumes inform subsequent category-specific evaluations. Because reliable application of these criteria requires quantitative measurement, visual estimation alone is insufficient [Schulz-Menger et al., 2020]. select_cardiac_phases(s)â (t ED ,t ES )identifies end-diastolic and end-systolic frames from the LV cavity volume curve. Accurate phase identification is critical because incorrect ED/ES selection introduces systematic volume errors that propagate to all downstream metrics [Schulz- Menger et al., 2020]. This tool automates the clinically recommended approach of selecting ED as maximum and ES as minimum LV cavity volume. Morphometric tool.compute_cardiac_metrics(s,v) âmderives advanced diagnostic fea- tures, each linked to a specific branch of the diagnostic Rule-Based Cardiac Phenotype Classifier: âąLV myocardial mass (V MYO Ă 1.05g/mL): a standard CMR metric for assessing ventricular hypertrophy [Arbelo et al., 2023]. âąMass-to-volume ratio (MVR): captures disproportionate myocardial thickening relative to chamber volume and supports discrimination between HCM and athleteâs heart [Czimbalmos et al., 2019]. âąMaximal end-diastolic wall thickness (EDWT): using an AHA segment-based representation and used to evaluate the encoded wall-thickness criteria. âąSeptal-to-lateral wall-thickness ratio (S:L): provides a quantitative descriptor of asymmetric septal hypertrophy and is used as one feature in the study-specific HCM-versus-HHD rule [Henry et al., 1973]. âąSphericity index (SI): LV geometry and is used in the study-specific Rule-Based Cardiac Phe- notype Classifier as a feature associated with spherical ventricular remodeling [Nakamori et al., 2017]. Wall motion tool.wall_motion_analysis(s)âwevaluates regional wall motion by comparing endocardial radial contraction and myocardial thickening between ED and ES frames. Each AHA segment is classified as normal, hypokinetic, or akinetic. This tool identifies focal versus diffuse wall-motion patterns. A focal territorial abnormality may support an ischemic phenotype, whereas diffuse hypokinesis may be more compatible with a nonischemic dilated phenotype; however, cine wall motion alone does not establish etiology or myocardial scar. 4 Training Pipeline We adopt a two-stage training paradigm inspired by recent agentic training approaches [Jiang et al., 2025, Shao et al., 2024], with modifications tailored to quantitative (rather than visual) tool use. Stage 1: Cold-start SFT. We fine-tune Qwen3-VL-8B-Instruct [Bai et al., 2025] with LoRA (rank 16,α = 32) on the training partition of the VQA benchmark using reference tool-use trajectories, whose displayed tool outputs are ground-truth-derived by default and may optionally be substituted with model-predicted measurements for trace realism; the boxed reference answer is always ground- truth-derived regardless. This stage trains the model to generate the tool-invocation syntax, follow the interleaved trace format, and associate question types with tools specified in the reference trajectories. We then augment this trajectory pool by rejection sampling: rollouts from the resulting policy are retained as additional SFT trajectories only if they yield a correct boxed answer, satisfy a tool-protocol adherence threshold, and, for classification items, are self-consistent across repeated samples. Stage 2: Agentic GRPO. We apply Group Relative Policy Optimization (GRPO) [Shao et al., 2024], which samples a group ofGresponseso i G i=1 from the current policyÏ Îž old for each promptqand optimizes: J GRPO (Ξ) =E q,o i " 1 G G X i=1 min r i (Ξ) Ë A i , clip r i (Ξ), 1âΔ, 1+Δ Ë A i â ÎČ D KL Ï Îž â„Ï ref # (1) wherer i (Ξ) = Ï Îž (o i | q)/Ï Îž old (o i | q)is the importance ratio, Ë A i = (S i â ÎŒ G )/Ï G is the group- normalized advantage computed from per-sample rewardsS i , andD KL Ï Îž â„Ï ref is a KL-divergence penalty (estimated with the unbiased low-variance estimator of Shao et al. [2024]) that regularizes the policy toward a fixed reference policyÏ ref , set to the Stage-1 SFT checkpoint and held frozen throughout Stage 2, with coefficient ÎČ. The composite reward is: 8 S = (1â w fmt )S ans + w fmt S format + w tool S tool + w live S live â w rep P rep (2) whereS ans â [0, 1]scores answer accuracy (tolerance-scaled for numeric tasks, exact match for classification),S format â [â0.5, 0]penalizes malformed outputs,S tool â [â0.1, +0.3]is conditional on answer correctness (tool bonuses are awarded only when the final answer is correct, which is intended to discourage indiscriminate tool invocation without corresponding answer accuracy),S live rewards successful live tool execution, andP rep penalizes degenerate repetition. Default weights are w fmt =0.15,w tool =0.12,w live =0.03,w rep =0.5; these are hand-set defaults rather than the product of a systematic hyperparameter sweep, and we did not evaluate sensitivity to them. 5 Experiments 5.1 Setup Datasets. We construct a multi-cohort cine CMR visual question answering benchmark from ACDC [Bernard et al., 2018], M&Ms [Campello et al., 2021], and M&Ms-2 [MartĂn-Isla et al., 2023]. The benchmark contains 11,531 question-answer pairs, including 7,020 training, 1,191 validation, and 3,320 test items across ten harmonized outcome categories. Each sample consists of a CMR volume, a benchmark question, a deterministic reference answer derived from the predefined decision rules, and annotations specifying the expected tool-use policy. For each source cohort, we construct patient-level partitions based on the official dataset splits before generating questions and paraphrases, and subsequently merge the partitions across cohorts. Thus, all images, questions, paraphrases, tool annotations, and derived measurements associated with a given patient are restricted to a single partition, preventing patient-level data leakage. The validation split is used for checkpoint selection and epoch-level monitoring during both training stages. The test split is used only for final evaluation and is never used for model selection or hyperparameter tuning. Metrics. A sampled response receives a score of1if its final boxed answer satisfies the itemâs reference-answer specification and0otherwise. Numerical answers are evaluated using prespecified item-level tolerances, while categorical and binary answers are evaluated using normalized exact matching. All reported rollouts execute predicted tool calls live through the tool runtime. We report pass@1 and pass@4 using the standard averaged pass@kestimator of Chen et al. [2021], pass@k :=E items 1â nâc k / n k , wherecis the number of correct responses amongnsampled rollouts for an item. Withn=4, pass@1 is the mean fraction of correct rollouts per item, while pass@4 indicates whether at least one of the four rollouts is correct, averaged across items. We further report two tool-use diagnostics (Figure 5a). The correct tool call rate is the fraction of gold tool-requiring items for which the predicted tool set exactly matches the gold tool set. The correct tool return rate is the fraction of executed gold-matched tool calls whose returned outputs agree with the ground truth within the prescribed tolerance. Implementation. We fine-tune Qwen3-VL-8B-Instruct on tool-integrated reasoning traces using LoRA with rankr=16, scaling factorα=32, and dropout0.05. SFT uses AdamW in bf16 with a 4,096-token context window. Starting from the resulting checkpoint, we perform tool-integrated GRPO using LoRA withr=64andα=128,G=4rollouts per prompt, a maximum response length of 1,024 tokens, and three training epochs. We use asymmetric clipping withΔ lo =0.2andΔ hi =0.3, and setÎČ=0.03in Eq. 1. The reward weights in Eq. 2 arew fmt =0.15,w tool =0.12,w live =0.03, and w rep =0.5. These values are fixed heuristically and are not tuned using validation or test performance. Training is performed on NVIDIA A100 and H100 GPUs. We report results for both the SFT checkpoint and the GRPO checkpoint. We compare against the unfine-tuned Qwen3-VL-8B-Instruct backbone [Bai et al., 2025] using the same CineMR tool- aware system prompt, so that this comparison isolates the effect of training from prompting. We additionally compare against the off-the-shelf medical VLMs LLaVA-Med v1.5 [Li et al., 2023] and MedGemma-4B-it [Sellergren et al., 2026]. SFT and GRPO are evaluated using four rollouts per item with temperature 0.7, from which pass@1 and pass@4 are computed. 9 Table 2: Held-out test performance by cohort (n = 3,320; ACDC: 696, M&Ms: 1,846, M&Ms-2: 778). All methods usek = 4rollouts with temperature 0.7. SFT and GRPO correspond to the matched pre- and post-RL models under the same live-tool protocol. GRPO consistently improves performance over SFT across all cohorts, yielding the best overall results. ACDCM&MsM&Ms-2Overall MethodP@1P@4MeanP@1P@4MeanP@1P@4MeanP@1P@4Mean LLaVA-Med v1.50.00.00.00.00.00.00.00.00.00.00.00.0 MedGemma-4B5.216.15.57.719.17.46.717.06.37.018.06.7 Qwen3-VL-8B1.34.91.41.44.21.41.94.61.41.54.51.4 SFT15.228.315.626.541.626.324.242.024.423.638.923.6 GRPO33.958.034.736.258.035.436.962.037.335.958.935.7 Table 3: Layer-wise performance of SFT (pre-RL) and GRPO (post-RL) with live tool execution, compared with the off-the-shelf MedGemma-4B-it and Qwen3-VL-8B baselines. Results are reported by cognitive layer on the held-out test set (n = 3,320). GRPO improves over SFT across most cognitive layers, with the largest gains observed in Layers 2â3. MedGemma-4BQwen3-VL-8BCineMR (SFT)CineMR (GRPO) SliceP@1P@4MeanP@1P@4MeanP@1P@4MeanP@1P@4Mean Layer 11.97.52.22.98.32.66.520.26.410.831.510.1 Layer 217.638.817.00.51.70.427.633.027.365.392.065.3 Layer 38.423.37.20.61.90.655.283.755.860.986.760.7 Layer 42.815.04.20.01.90.5 50.578.552.145.878.550.0 Layer 50.00.00.00.00.00.012.933.312.613.332.513.0 Layer 60.01.50.40.01.50.437.950.035.637.937.937.1 All layers7.018.06.71.54.51.423.638.923.635.958.935.7 5.2 Results Table 2 presents the main results. GRPO raises pass@1 from 23.6% (cold-start SFT) to 35.9% overall, with the largest single-cohort gain on ACDC (15.2%â33.9%) and consistent gains on M&Ms (26.5%â36.2%) and M&Ms-2 (24.2%â36.9%). Under the samek=4-averaged pass@1 protocol used throughout, LLaVA-Med produces no valid diagnoses (0.0%), MedGemma reaches 7.0%, and the un-fine-tuned Qwen3-VL-8B backbone reaches 1.5%; CineMRâs pass@1 is roughly24Ăhigher than the un-fine-tuned backbone it is trained from, and more than5Ăhigher than the strongest off-the-shelf baseline. 5.3 Ablations Table 3 breaks down the main results by cognitive layer. On the clinical-criteria layers (L3âL4), where a measured value must be evaluated against a predefined clinical criterion, GRPO reaches 58.7% pass@1, compared with 54.5% for SFT and at most 3.0% for either off-the-shelf baseline. This large gap suggests that explicit tool-integrated quantitative reasoning is particularly important when assessment depends on measured evidence rather than qualitative visual appearance alone. On the full-diagnosis layers (L5âL6), GRPO (18.6%) only marginally improves over SFT (18.3%), consistent with the compounding difficulty of multi-step reasoning, where errors in upstream segmentation or measurement can propagate through subsequent decision steps. Thus, CineMRâs gains are not uniform across visual question answering tasks, but are concentrated on tasks that require converting cine MRI into explicit quantitative evidence and evaluating that evidence against clinical criteria. Figure 5a further shows that GRPO primarily improves the precision of tool selection. GRPO predicts the exact gold tool set on 99.8% of gold tool-requiring items, compared with 78.9% for SFT. Figure 5b breaks this down further: across the four volumetric and ejection-fraction question types, the live measure_volumetoolâs own returned value is correct substantially more often (28â34%) than the modelâs final boxed answer (4â11%), indicating that part of this gap reflects how the model integrates a returned measurement rather than the measurementâs own accuracy. 10 Correct Tool Call Rate 0 25 50 75 100 % of gold-required items 78.9 99.8 +20.9 SFT (pre-RL)GRPO (post-RL) (a) Tool call rate. LVEFLV volumes RV volumes RVEF 0 10 20 30 40 % within tolerance of reference 8.8 11.2 7.5 4.2 30.0 34.2 27.927.9 +21.2 +23.0 +20.4 +23.7 Model's answerLive tool value (b) Model answer vs. live tool value. All tools (baseline) w/o computation tools w/o all tools 0 10 20 30 40 pass@1 (%) 35.9 34.8 27.9 â8.0 pts All tools (baseline)w/o computation toolsw/o all tools (c) Pass@1 with tools withheld. Figure 5: Tool-use diagnostics and tool-withholding ablation. (a) Correct tool-call rate for SFT and GRPO on gold tool-requiring items. (b) GRPO accuracy of direct model predictions versus live measure_volumeoutputs on ventricular volume and ejection-fraction questions. (c) GRPO pass@1 on the held-out test set (n = 3,320,k = 4) with full tool access, without computation tools, and without any tools; error bars show 95% bootstrap confidence intervals. GRPO raises correct tool invocation from 78.9% to 99.8%; live tool outputs improve measurement accuracy by 20.4â23.7%, and removing all tools lowers pass@1 from 35.9% to 27.9%. We next test how much these quantitative outputs contribute to final performance. Withholding the three metric-producing tools at inference time, while retaining segmentation, wall-motion analysis, and visual zoom, reduces pass@1 from 35.9% to 34.8% (Figure 5c). Without explicit measurements, the model must instead rely on visual cues and learned priors for questions governed by numerical criteria. Removing all tools, including segmentation, further reduces pass@1 to 27.9%. The 95% bootstrap confidence interval for this setting does not overlap with that of the full-tool configuration, indicating that the degradation is unlikely to be attributable to sampling variation alone. Together, these results show that externally computed image-derived evidence contributes information beyond what the VLM recovers from visual reasoning alone. 6 Conclusion We presented CineMR, a tool-augmented vision-language framework for quantitative cine CMR assessment. CineMR combines visual interpretation with dedicated tools for segmentation, phase selection, ventricular volumetry, morphometric measurement, and regional wall-motion analysis, and maps the resulting evidence to benchmark categories using an explicitly encoded Rule-Based Cardiac Phenotype Classifier. On the constructed multi-cohort benchmark, supervised fine-tuning and reinforcement learning substantially improved performance relative to the unfine-tuned backbone, while access to quantitative tools provided an additional, more modest improvement over the tool- ablated configuration. These findings demonstrate the potential of integrating vision-language models with explicit quantitative computation for traceable assessment of cine CMR examinations. However, the benchmark categories and study-specific rules do not constitute complete clinical diagnoses, and the framework has not been validated for autonomous clinical use. Future work should evaluate external generalizability, measurement robustness, calibration, and clinical utility across institutions, scanner vendors, patient populations, and CMR acquisition protocols. Acknowledgments and Funding This work utilized research computing resources supported by the Advanced Cyberinfrastructure Coor- dination Ecosystem: Services & Support (ACCESS) program under allocation ACCESS-CIS240958, including resources provided by the Pittsburgh Supercomputing Center Bridges-2 system. We also thank the AI-MIND Lab for providing GPU resources used in this work. The authors declare no competing interests. 11 References Steffen Erhard Petersen, Paul M. Matthews, Jane M. Francis, Matthew D. Robson, Filip Zem- rak, Redha Boubertakh, Alistair A. Young, Sarah Hudson, Peter J. Weale, Steve Garratt, Rory Collins, Stefan K. Piechnik, and Stefan Neubauer. UK Biobankâs cardiovascular mag- netic resonance protocol. Journal of Cardiovascular Magnetic Resonance, 18, 2016. URL https://api.semanticscholar.org/CorpusID:13942043. Jeanette Schulz-Menger, David A. Bluemke, Jens Bremerich, Scott D. Flamm, Mark A. Fogel, Matthias G. Friedrich, Raymond J. Kim, Florian von Knobelsdorff-Brenkenhoff, Christopher M. Kramer, Dudley J. Pennell, Sven Plein, and Eike Nagel. Standardized image interpretation and post- processing in cardiovascular magnetic resonance â 2020 update: Society for Cardiovascular Mag- netic Resonance (SCMR): Board of Trustees Task Force on Standardized Post-Processing. Journal of Cardiovascular Magnetic Resonance, 22(1):19, 2020. doi: 10.1186/s12968-020-00610-6. Elena Arbelo, Alexandros Protonotarios, Juan RamĂłn Gimeno, Eloisa Arbustini, Roberto Barriales- Villa, Cristina Basso, Connie R. Bezzina, Elena Biagini, Nico A Blom, Rudolf A. de Boer, Tim De Winter, Perry M. Elliott, Marcus D Flather, Pablo GarcĂa-PavĂa, Kristina Hermann Haugaa, Jodie Ingles, Ruxandra Oana Jurcut, Sabine Klaassen, Giuseppe Limongelli, Bart L. Loeys, Jens Mogensen, Iacopo Olivotto, Antonis Pantazis, Sanjay Sharma, Jan Peter van Tintelen, James S. Ware, Juan Pablo Kaski, Philippe Charron, Massimo Imazio, Magdy Abdelhamid, Victor Aboyans, Michael Arad, Folkert W. Asselbergs, Riccardo Asteggiano, Zofia Teresa Bili Ì nska, Damien Bonnet, Henning Bundgaard, Nuno Cardim, Jelena Ë Celutkien Ì e, Maja Ë CikeĆĄ, Gaetano Maria de Ferrari, Veronica Dusi, Volkmar Falk, Laurent Fauchier, Estelle Gandjbakhch, Tiina Heliö, Konstantinos C. Koskinas, Dipak Kotecha, Ulf Landmesser, George Lazaros, Basil S. Lewis, A. M. Linhart, Maja-Lisa LĂžchen, Benjamin Meder, Richard Mindham, James C Moon, Jens Cosedis Nielsen, Steffen E Petersen, Eva Bossano Prescott, Mary N. Sheppard, Gianfranco Sinagra, Marta Sitges, Jacob Tfelt-Hansen, Rhian M. Touyz, Rogier J. A. Veltrop, Josef Veselka, Karim Wahbi, Arthur A. M. Wilde, Katja Zeppenfeld, Brahim Kichou, Hamayak S. Sisakian, Daniel Scherr, Bernhard L. Gerber, Alen DĆŸubur, Mariana Gospodinova, Ivo Planinc, Hera Heracleous Moustra, David ZemĂĄnek, Morten Steen Kvistholm Jensen, Ahmad Samir, Kairit Palm, Eric Schulze- Bahr, Vlachopoulos Haralambos, RĂłbert Sepp, Berglind HlĂn AĂ°alsteinsdĂłttir, Deirdre Ward, Miry Blich, Afrim Poniku, Olga Lunegova, Ainars Rudzitis, Roland Kassab, J Ì urat Ì e Barysien Ì e, Steve Huijnen, T. P. De Felice, Eleonora Vataman, Nikola Pavlovi Ì c, Nawal Doghmi, Folkert W. Asselbergs, Elizabeta Srbinovska Kostovska, Vibeke M Almaas, El Ì zbieta Katarzyna Biernacka, Dulce Brito, Monica Ros , ca, Marco Zavatta, Arsen Risti Ì c, Eva Goncalvesova, MatjaĆŸ Ć inkovec, Victoria Cañadas-Godoy, Pyotr G. Platonov, Ardan Muammer Saguner, Ahmad Al Saadi, Ikram Kammoun, Ahmet Ăelik, E. G. Nesukay, Timur A. Abdullaev, Stefan James, Elena Arbelo, Colin Baigent, Michael A Borger, Sergio Buccheri, Borja Ibåñez, Lars KĂžber, Konstantinos C. Koskinas, John Mcevoy, Borislava Mihaylova, Lis Neubeck, AgnĂšs Pasquet, Amina Rakisheva, Bianca Rocca, Xavier RossellĂł, Ilonca Vaartjes, Christiaan J M Vrints, and Adam Witkowski. 2023 ESC Guidelines for the management of cardiomyopathies. European Heart Journal, 2023. URL https://api.semanticscholar.org/CorpusID:261121505. Yuxiang Lai, Jike Zhong, Ming Li, Shitian Zhao, and Xiaofen Yang. Med-R1: Reinforcement Learn- ing for Generalizable Medical Reasoning in Vision-Language Models. IEEE Transactions on Med- ical Imaging, P, 2025. URL https://api.semanticscholar.org/CorpusID:277104458. Jiazhen Pan, Che Liu, Junde Wu, Fenglin Liu, Jiayuan Zhu, Hongwei Bran Li, Chen Chen, Ouyang Cheng, and Daniel Rueckert. MedVLM-R1: Incentivizing Medical Reasoning Capability of Vision-Language Models (VLMs) via Reinforcement Learning. In International Conference on Medical Image Computing and Computer-Assisted Intervention, 2025. URLhttps://api. semanticscholar.org/CorpusID:276647067. Yan-Cheng Su, Tian-Xin Li, Jiyao Liu, Chenglong Ma, Junzhi Ning, Cheng Tang, Sibo Ju, Jin Ye, Pengcheng Chen, Ming Hu, Shixiang Tang, Lihao Liu, Bin Fu, Wenqi Shao, Xi- aowei Hu, Xiangwen Liao, Yuanfeng Ji, and Junjun He. GMAI-VL-R1: Harnessing Rein- forcement Learning for Multimodal Medical Reasoning. ArXiv, abs/2504.01886, 2025. URL https://api.semanticscholar.org/CorpusID:277502243. 12 Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. LLaVA-Med: Training a Large Language-and- Vision Assistant for Biomedicine in One Day. ArXiv, abs/2306.00890, 2023. URLhttps: //api.semanticscholar.org/CorpusID:258999820. Haoran Sun, Yankai Jiang, Wenjie Lou, Yujie Zhang, Wenjie Li, Lilong Wang, Mianxin Liu, Lei Liu, and Xiaosong Wang. Chiron-o1: Igniting Multimodal Large Language Models towards Generalizable Medical Reasoning via Mentor-Intern Collaborative Search. In Advances in Neural Information Processing Systems (NeurIPS), 2025. Steffen Erhard Petersen, Nay Linn Aung, Mihir M. Sanghvi, Filip Zemrak, Kenneth Fung, JosĂ© Miguel Paiva, Jane M. Francis, Mohammed Yunus Khanji, Elena Lukaschuk, Aaron Mark Lee, Valentina Carapella, Young Jin Kim, Paul Leeson, Stefan K. Piechnik, and Stefan Neubauer. Reference ranges for cardiac structure and function using cardiovascular magnetic resonance (CMR) in Caucasians from the UK Biobank population cohort. Journal of Cardiovascular Magnetic Resonance, 19, 2017. URL https://api.semanticscholar.org/CorpusID:4240471. Binxu Li, Tian Yan, Yuanting Pan, Zhe Xu, Jie Luo, Ruiyang Ji, Shilong Liu, Haoyu Dong, Zihao Lin, and Yixin Wang. MMedAgent: Learning to Use Medical Tools with Multi-modal Agent. ArXiv, abs/2407.02483, 2024. URL https://api.semanticscholar.org/CorpusID:270877969. Nima Fathi, Amar Kumar, and Tal Arbel. AURA: A Multi-Modal Medical Agent for Understanding, Reasoning & Annotation. ArXiv, abs/2507.16940, 2025. URLhttps://api.semanticscholar. org/CorpusID:280224709. Ziyue Wang, Junde Wu, Linghan Cai, Chang Han Low, Xihong Yang, Qiaxuan Li, and Yueming Jin. MedAgent-Pro: Towards Evidence-based Multi-modal Medical Diagnosis via Reasoning Agentic Workflow. In International Conference on Learning Representations (ICLR), 2026. Yankai Jiang, Yujie Zhang, Peng Zhang, Yichen Li, Jintai Chen, Xiaoming Shi, and Shihui Zhen. Incentivizing Tool-augmented Thinking with Images for Medical Image Analysis. ArXiv, abs/2512.14157, 2025. URL https://api.semanticscholar.org/CorpusID:283909495. Yunguan Fu, Wenjia Bai, Weixi Yi, Charlotte Manisty, Anish N. Bhuva, Thomas A. Treibel, James C. Moon, Matthew J. Clarkson, Rhodri H. Davies, and Yipeng Hu. CineMA: A versatile foundation model for cine cardiac magnetic resonance image analysis tasks. arXiv preprint arXiv:2506.00679, 2025. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Jun-Mei Song, Mingchuan Zhang, Y. K. Li, Yu Wu, and Daya Guo. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. ArXiv, abs/2402.03300, 2024. URLhttps://api.semanticscholar.org/ CorpusID:267412607. Olivier Bernard, Alain Lalande, ClĂ©ment Zotti, FrĂ©dĂ©ric Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel A. gonzĂĄlez Ballester, Gerard SanromĂĄ, Sandy Napel, Steffen Erhard Petersen, Georgios Tziritas, Elias Grinias, Mahendra Khened, Varghese Alex Kollerathu, Ganapathy Krishnamurthi, Marc-Michel RohĂ©, Xavier Pennec, Maxime Sermesant, Fabian Isensee, Paul F. JĂ€ger, Klaus Hermann Maier-Hein, Peter M. Full, Ivo Wolf, Sandy Engelhardt, Christian F. Baumgartner, Lisa M. Koch, Jelmer M. Wolterink, Ivana IĆĄgum, Yeonggul Jang, Yoonmi Hong, Jay Patravali, Shubham Jain, Olivier Humbert, and Pierre-Marc Jodoin. Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis: Is the Problem Solved? IEEE Transactions on Medical Imaging, 37:2514â2525, 2018. URL https://api.semanticscholar.org/CorpusID:51610194. VĂctor M. Campello, Polyxeni Gkontra, Cristian Izquierdo, Carlos MartĂn-Isla, Alireza Sojoudi, Peter M. Full, Klaus Hermann Maier-Hein, Yao Zhang, Zhiqiang He, Jun Ma, Mario Parreño, Alberto Albiol, Fanwei Kong, Shawn C. Shadden, Jorge Corral Acero, Vaanathi Sundaresan, Mina Saber, Mustafa Elattar, Hongwei Li, Bjoern H Menze, Firas Khader, Christoph Haarburger, Cian M. Scannell, Mitko Veta, Adam Carscadden, K. Punithakumar, Xiao Liu, Sotirios A. Tsaftaris, Xiaoqiong Huang, Xin Yang, Lei Li, Xiahai Zhuang, David ViladĂ©s, Martin Descalzo, Andrea Guala, Lucia La Mura, Matthias G. Friedrich, Ria Garg, Julie Lebel, Filipe Henriques, Mahir 13 Karakas, Ersin Cavus, Steffen Erhard Petersen, Sergio Escalera, S. SeguĂ, JosĂ© F. RodrĂguez Palo- mares, and Karim Lekadir. Multi-Centre, Multi-Vendor and Multi-Disease Cardiac Segmentation: The M&Ms Challenge. IEEE Transactions on Medical Imaging, 40:3543â3554, 2021. URL https://api.semanticscholar.org/CorpusID:235471536. Carlos MartĂn-Isla, VĂctor M. Campello, Cristian Izquierdo, Kaisar Kushibar, Carla Sendra-Balcells, Polyxeni Gkontra, Alireza Sojoudi, Mitchell J. Fulton, Tewodros Weldebirhan Arega, K. Punithaku- mar, Lei Li, Xiaowu Sun, Yasmina Al Khalil, Ding Liu, Sana Jabbar, Sandro QueirĂłs, Francesco Galati, Moona Mazher, Zheyao Gao, Marcel Beetz, Lennart Tautz, Christoforos Galazis, Marta Varela, Markus Hullebrand, Vicente Grau, Xiahai Zhuang, Domenec Puig, Maria A. Zuluaga, Hassan Mohy ud Din, Dimitris N. Metaxas, Marcel M. Breeuwer, Rob J. van der Geest, Michelle L. Noga, StĂ©phanie Bricq, Mark E. Rentschler, Andrea Guala, Steffen E. Petersen, Sergio Escalera, JosĂ© F. RodrĂguez Palomares, and Karim Lekadir. Deep Learning Segmentation of the Right Ventri- cle in Cardiac MRI: The M&Ms Challenge. IEEE Journal of Biomedical and Health Informatics, 27:3302â3313, 2023. URL https://api.semanticscholar.org/CorpusID:258178794. Vishwesh Nath, Wenqi Li, Dong Yang, Andriy Myronenko, Mingxin Zheng, Yao Lu, Zhijian Liu, Hongxu Yin, Yee Man Law, Yucheng Tang, Pengfei Guo, Can Zhao, Ziyue Xu, Yufan He, Greg Heinrich, Stephen R. Aylward, Marc Edgar, Michael Zephyr, Pavlo Molchanov, Baris I Turkbey, Holger R. Roth, and Daguang Xu. VILA-M3: Enhancing Vision-Language Models with Medical Expert Knowledge. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14788â14798, 2025. URLhttps://api.semanticscholar.org/CorpusID: 274150020. Qiao Jin, Zhizheng Wang, Yifan Yang, Qingqing Zhu, Donald Wright, Thomas Huang, W. John Wilbur, Zhe He, Andrew Taylor, Qingyu Chen, and Zhiyong Lu. AgentMD: Empowering Language Agents for Risk Prediction with Large-Scale Clinical Tool Learning. ArXiv, 2024. URLhttps: //api.semanticscholar.org/CorpusID:267759870. Fabian Isensee, Paul F. Jaeger, Simon A. A. Kohl, Jens Petersen, and Klaus Hermann Maier-Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18:203 â 211, 2021. URLhttps://api.semanticscholar.org/CorpusID: 227947847. Bram Ruijsink, Esther Puyol-AntĂłn, Ilkay Oksuz, Matthew Sinclair, Wenjia Bai, Julia Anne Schnabel, Reza Razavi, and Andrew P. King. Fully Automated, Quality-Controlled Cardiac Analysis From CMR: Validation and Large-Scale Application to Characterize Cardiac Function. JACC: Cardiovascular Imaging, 13:684 â 695, 2020. URLhttps://api.semanticscholar.org/ CorpusID:198134360. Irem Cetin, Zahra Raisi-Estabragh, Steffen Erhard Petersen, Sandy Napel, Stefan K. Piechnik, Stefan Neubauer, Miguel A. gonzĂĄlez Ballester, Oscar Camara, and Karim Lekadir. Radiomics Signatures of Cardiovascular Risk Factors in Cardiac MRI: Results From the UK Biobank. Frontiers in Cardiovascular Medicine, 7, 2020. URLhttps://api.semanticscholar.org/CorpusID: 226222975. Frank I. Marcus, William J. McKenna, Duane L. Sherrill, Cristina Basso, Barbara Bauce, David A Bluemke, Hugh Calkins, Domenico Corrado, Moniek G.P.J. Cox, James P. Daubert, Guy Hugues Fontaine, Kathleen Gear, Richard N W Hauer, Andrea Nava, Michael H. Picard, Nikos Protono- tarios, Jeffrey E. Saffitz, Danita M. Yoerger Sanborn, Jonathan S. Steinberg, Harikrishna Tan- dri, Gaetano Thiene, Jeffrey A. Towbin, Adalena Tsatsopoulou, Thomas Wichter, and Wo- jciech Zareba. Diagnosis of Arrhythmogenic Right Ventricular Cardiomyopathy/Dysplasia: Proposed Modification of the Task Force Criteria. Circulation, 121:1533â1541, 2010. URL https://api.semanticscholar.org/CorpusID:15324668. Jens Vogel-Claussen, John Paul Finn, Antoinette S Gomes, Gregory W. Hundley, Michael Jerosch- Herold, Gregory D. N. Pearson, Shantanu Sinha, Joao A. C. Lima, and David A. Bluemke. Left Ventricular Papillary Muscle Mass: Relationship to Left Ventricular Mass and Volumes by Magnetic Resonance Imaging. Journal of Computer Assisted Tomography, 30:426â432, 2006. URL https://api.semanticscholar.org/CorpusID:39925049. 14 D.H. Frank Gommans, Jeannette Bakker, Gilbert E Cramer, Freek W. A. Verheugt, Marc A. Brouwer, and Marcel J M Kofflard. Impact of the papillary muscles on cardiac magnetic resonance image analysis of important left ventricular parameters in hypertrophic cardiomyopathy. Netherlands Heart Journal, 24:326 â 331, 2016. URLhttps://api.semanticscholar.org/CorpusID: 215679534. Steve R. Ommen, Seema Mital, Michael A. Burke, Sharlene M. Day, Anita Deswal, Perry M. El- liott, Lauren L. Evanovich, Judy W. Hung, Jose A. Joglar, Paul F. Kantor, Carey D. Kimmelstiel, Michelle M. Kittleson, Mark S. Link, Martin S. Maron, Matthew W. Martinez, Christina Y. Miyake, Hartzell V. Schaff, Christopher Semsarian, and Paul Sorajja. 2020 AHA/ACC Guideline for the Di- agnosis and Treatment of Patients With Hypertrophic Cardiomyopathy: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guide- lines. Circulation, 2020. URL https://api.semanticscholar.org/CorpusID:227078319. Csilla Czimbalmos, Ibolya CsĂ©cs, Attila TĂłth, Orsolya Kiss, Ferenc Imre Suhai, NĂłra SydĂł, ZsĂłfia Dohy, Astrid Apor, BĂ©la Merkely, and Hajnalka VĂĄgĂł. The demanding grey zone: Sport indices by cardiac magnetic resonance imaging differentiate hypertrophic cardiomyopathy from athleteâs heart. PLoS ONE, 14, 2019. URL https://api.semanticscholar.org/CorpusID:67857631. Walter L. Henry, Chester E. Clark, and Stephen E. Epstein. Asymmetric Septal Hypertrophy: Echocar- diographic Identification of the Pathognomonic Anatomic Abnormality of IHSS. Circulation, 47: 225â233, 1973. URL https://api.semanticscholar.org/CorpusID:6722097. Shiro Nakamori, Haisam Ismail, Long H. Ngo, Warren J. Manning, and Reza Vatani Nezafat. Left ventricular geometry predicts ventricular tachyarrhythmia in patients with left ventricular systolic dysfunction: a comprehensive cardiovascular magnetic resonance study. Journal of Cardiovascular Magnetic Resonance, 19, 2017. URLhttps://api.semanticscholar.org/ CorpusID:28351378. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-VL Technical Report. arXiv preprint arXiv:2511.21631, 2025. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating Large Language Models Trained on Code. arXiv preprint arXiv:2107.03374, 2021. Andrew Sellergren, Chufan Gao, Fereshteh Mahvar, Timo Kohlberger, Fayaz Jamil, Madeleine Traverse, Alberto Tono, Bashir Sadjad, Lin Yang, Charles Lau, et al. MedGemma 1.5 Technical Report. arXiv preprint arXiv:2604.05081, 2026. 15 A Acronym Glossary Table 4 collects every abbreviation used in the main text and appendix. Each is also spelled out at first use in the running text; this table is provided as a single point of reference. Table 4: Acronym glossary. AcronymDefinition LV / RVLeft / right ventricle (ventricular) LVEF / RVEFLeft / right ventricular ejection fraction LVEDV / LVESVLV end-diastolic / end-systolic volume RVEDV / RVESVRV end-diastolic / end-systolic volume EDWTEnd-diastolic wall thickness S:L ratioSeptal-to-lateral wall thickness ratio SISphericity index MVRMass-to-volume ratio (LV mass / LVEDV) WMAWall motion abnormality BSABody surface area HCMHypertrophic cardiomyopathy HHDHypertensive heart disease DCMDilated cardiomyopathy ARVCArrhythmogenic right ventricular cardiomyopathy MINFMyocardial infarction (ischemic scar) IHDIschemic heart disease AHSAthleteâs heart syndrome LVNCLeft ventricular noncompaction NORNormal (no structural/functional abnormality) rTFC(2010) Revised Task Force Criteria for ARVC ESC / AHAEuropean Society of Cardiology / American Heart Association SCMRSociety for Cardiovascular Magnetic Resonance CMRCardiac magnetic resonance B Evaluation Question Templates Section 3 (âCognitive layer hierarchyâ) defines the six layers and gives one illustrative example question per layer. Table 5 lists the complete set of 19 question templates used to construct the ACDC / M&Ms / M&Ms-2 benchmark (Section 5.1), each mapped to its cognitive layer, tool requirement, and the tools it invokes when tools are used. Each templateâs literal question text below is the canonical form; the released dataset additionally includes two LLM-generated paraphrase variants per base question (rephrasing, context enrichment, and specificity variation) for lexical diversity, so a given evaluation item may appear as a paraphrase of, rather than verbatim match to, the text shown here. The canonical templates, their answer-derivation functions, and applicability conditions are implemented in the released code, which is the authoritative source for exact question text, answer formats, and clinical citations. 16 Table 5: All 19 evaluation question templates, by cognitive layer. Tools: whether tool use is required (â) or optional (â) for that template; a template may override its layerâs default (e.g. L2_is_asymmetric requires tools because the septal:lateral ratio must be computed). Template IDLayer Tools Question L1_lvefL1âWhat is the left ventricular ejection fraction (LVEF) for this patient? Report as a percentage and indicate whether it is normal, mildly reduced, reduced, or severely reduced. L1_lvedvL1âReport the left ventricular end-diastolic volume (LVEDV) and end-systolic volume (LVESV) for this cardiac MRI study, including BSA-indexed values if anthropometric data are available. L1_rvedvL1âWhat are the right ventricular end-diastolic volume (RVEDV), end-systolic volume (RVESV), and RV ejection fraction (RVEF) for this patient? L1_lv_massL1âCalculate the left ventricular myocardial mass from this cardiac MRI segmentation. Report the absolute LV mass in grams. L1_rvefL1âWhat is the right ventricular ejection fraction (RVEF) for this patient? Is it within normal limits? L1_max_edwtL1âWhat is the maximum end-diastolic wall thickness (max EDWT) of the left ventricle in this cardiac MRI, and does it exceed the threshold for hypertrophy? L2_is_dilatedL2â Is the left ventricle dilated in this cardiac MRI? Justify your answer using volumetric measurements. L2_is_hypertrophiedL2âDoes this patient demonstrate left ventricular hypertrophy? Report the maximum wall thickness and indicate whether it is within normal limits. L2_is_asymmetricL2âIs the left ventricular hypertrophy asymmetric? Compute the septal-to-lateral wall thickness ratio. L2_has_wmaL2âDoes this cardiac MRI demonstrate regional wall motion abnormalities? If so, which AHA segments are affected and which coronary territory is involved? L3_hcm_thresholdL3âDoes the maximum end-diastolic wall thickness meet the ESC 2023 criterion for hypertrophic cardiomyopathy (â„15 m for sporadic cases)? Also assess whether theâ„13 m familial HCM criterion is met. L3_arvc_criteriaL3âDo the RV volumetric measurements meet the 2010 Revised Task Force Criteria (rTFC) for ARVC? Assess both major and minor CMR criteria for RVEDVi and RVEF. L3_minf_scarL3âIn the segment(s) with regional wall motion abnormality, does the end-diastolic wall thickness meet the criterion for probable transmural scar (â€5.5 m)? Assess the likelihood of chronic myocardial infarction versus viable hibernating myocardium. L3_dcm_siL3âIs the left ventricular sphericity index elevated above the threshold for pathological sphericization (>0.57)? Compute the sphericity index from the short-axis stack. L4_hcm_vs_hhdL4âThis patient has left ventricular hypertrophy with max EDWTâ„13 m. Is the hypertrophy pattern more consistent with hypertrophic cardiomyopathy (HCM) or hypertensive heart disease (HHD)? Use the S:L ratio, septal convexity, and mass-to-volume ratio in your assessment. L4_dcm_vs_ahsL4âThe left ventricle is dilated with reduced or borderline ejection fraction and no focal wall motion abnormality. Distinguish between dilated cardiomyopathy (DCM) and athleteâs heart syndrome (AHS) using the sphericity index, mass-to-volume ratio, and biventricular volume balance. L5_full_diagnosisL5âAnalyze this cardiac cine MRI and provide a comprehensive diagnostic assessment. Systematically evaluate the left and right ventricular morphology and function, identify any regional wall motion abnormalities, and arrive at the most likely diagnosis. Use the hypothesis-verification approach: state your initial phenotypic hypothesis, then verify with quantitative measurements. L6_differential_ dcm_minf L6â This patient has reduced left ventricular ejection fraction and ventricular dilation. The key differential is dilated cardiomyopathy (DCM) versus ischemic heart disease with prior myocardial infarction (MINF). What quantitative features allow you to distinguish between these two diagnoses? L6_differential_ hcm_athlete L6âThis patient is a competitive athlete with increased left ventricular wall thickness (max EDWT in the 13â15 m grey zone). Distinguish between hypertrophic cardiomyopathy (HCM) and physiological adaptation (athleteâs heart) using mass-to-volume ratio, septal shape, and cavity dimensions. 17 C Limitations and Broader Impact Segmentation dependency.CineMR relies on a dedicated short-axis segmentation stage to derive quantitative cardiac metrics, including ventricular volumes, ejection fractions, myocardial mass, and wall thickness. Because downstream benchmark classification depends on these segmentation- derived measurements, segmentation errors can propagate to the final predicted category. In our error analysis, under- or over-segmentation of the left ventricular cavity was a frequent failure mode in L5 full-assessment tasks and could direct the Rule-Based Cardiac Phenotype Classifier toward an incorrect branch even when the subsequent reasoning process was otherwise consistent with the extracted measurements. Toward segmentation-free quantitative reasoning. As multimodal foundation models and end- to-end learning methods advance, future systems may estimate quantitative cardiac measurements directly from imaging data within an integrated reasoning framework, potentially bypassing explicit segmentation. Such approaches may reduce segmentation-related error propagation; however, their measurement accuracy, robustness, and generalizability across acquisition protocols and scanner vendors will require dedicated validation. Broader impact. By recording tool calls, quantitative outputs, and criterion-based classification paths, CineMR provides more inspectable evidence traces than an end-to-end prediction alone. With prospective external validation, the framework could potentially support quantitative cine CMR assessment in settings where specialist interpretation is limited. However, CineMR is not intended for autonomous clinical use. Its outputs should be reviewed by qualified clinicians, and overreliance on automated results particularly when upstream segmentation is inaccurate could contribute to assessment errors if the system were deployed without appropriate validation, quality control, and safeguards. Additional limitations.This retrospective study uses public research datasets and primarily evalu- ates short-axis cine CMR. Several benchmark categories incorporate study-specific rules or thresholds and therefore should not be interpreted as complete clinical diagnoses. Future work should include prospective external validation across institutions, scanner vendors, and patient populations and assess calibration, subgroup performance, and robustness to variations in image quality and acquisition protocols. 18