CineMR: AI Model for Cardiac MRI That Calls Its Own Tools
CineMR combines vision-language AI with specialized cardiac analysis tools, achieving 35.9% accuracy on quantitative heart MRI tasks where base models score just 1.5%.
👁 112 reads
Cardiac magnetic resonance imaging (CMR), including cine imaging, is considered the gold standard for non-invasive assessment of heart structure and ventricular function. Interpreting these scans requires combining visual assessment with quantitative measurements: ventricular volumes, ejection fraction, myocardial mass, wall thickness, and regional wall motion. The problem is that existing medical vision-language models (VLMs) cannot reliably extract such data from multidimensional cine images without analytical tool support.
What CineMR Is and How It Works
CineMR is a VLM extended with the ability to call external tools for cardiac image analysis. The model can use tools for segmentation, phase selection, volumetry, morphometry, and regional wall motion analysis. The results from these tools are woven directly into the model’s reasoning process - hence the name “tool-integrated reasoning”. The authors also built their own benchmark for evaluating models: a multi-center set of questions and answers covering quantitative measurement extraction, multi-class diagnosis, and differential diagnosis.
How the Model Was Trained
Training proceeded in two stages. First, they applied supervised fine-tuning (SFT) on traces of tool interactions. Then they used Group Relative Policy Optimization (GRPO) with rewards for conditional tool use. After the SFT stage, tool-calling accuracy was 78.9%, and after GRPO it rose to 99.8%.
Benchmark Results
On the multi-center cine CMR benchmark, CineMR achieved 35.9% pass@1 (correct answer on first attempt) and 58.9% pass@4 (correct answer within four attempts). For comparison, the base Qwen3-VL-8B model scored 1.5% pass@1, LLaVA-Med v1.5 scored 0.0% pass@1, and MedGemma-4B scored 7.0% pass@1.
The authors also tested how much the tools themselves contribute. Live tool results improved ventricular measurement accuracy by 20.4-23.7% compared to direct model predictions. Removing all tools dropped pass@1 from 35.9% to 27.9%.
Resource Availability
The authors report that code, benchmark resources, and model weights are publicly available at the address provided in the paper. The paper is dated October 1, 2026, and was submitted by Kunyang Li and eight co-authors.
The document contains no information about whether the model was tested in clinical settings, what deployment plans exist, or whether it has undergone any regulatory validation.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →What This Means
In our view, the most important takeaway from this paper is that a large vision model alone - even a good one - is useless for quantitative cardiac diagnostics: the 1.5% effectiveness of the base model is essentially random performance. Only combining the model with specialized analytical tools produces something useful. This suggests that in medical imaging, the path to AI does not run through ever-larger general models, but through tight integration of language models with domain-specific computational tools. The 35.9% pass@1 result is still far from clinical utility, but the authors don’t claim the model is ready for deployment - they present it as a promising assistive approach, which we consider an honest framing.
Source: arXiv cs.AI: CineMR: Tool-Integrated Vision-Language Reasoning for Quantitative Cardiac MRI Assessment (document dated 2026-10-02). This summary is based on the full text of the paper, not reports from other media.
More on this from QUANTSLANT
news·Anthropic Researcher Quits, Warns Self-Improving AI Could ‘Kill Us All’