Research map / Vision-language and multimodal / Multimodal LLMs

CMR-RD: Long-Tailed Adaptive VLM for Explainable CMR Diagnosis

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CMR-RD, a Gemma-4B-based VLM for explainable cardiac MRI diagnosis that uses two-stage training (medical alignment with cold-start CoT, then GPPO RL with Thompson-sampling class balancing and an IoU lesion-grounding reward).

Problem

Automated CMR interpretation lacks verifiable reasoning chains and VLMs underperform on rare cardiac diseases under long-tailed class distributions.

Approach

Group-Phased Policy Optimization couples GRPO-style RL with per-stage Thompson-sampling class quotas (Beta posteriors over per-class accuracy) so that harder or rarer classes are sampled more often, plus an IoU-based lesion-grounding reward to reduce hallucinated lesion descriptions.

Key results

• CMR-RD reaches ACC 0.610, F1 0.639, AUC 0.703 overall in the ablation (S1+S2), versus 0.212 with neither stage.
• Per-class table reports CMR-RD ACC 0.641 on HCM and 0.574 on CAM, outperforming listed general and medical VLMs, which score near 0 on several rare classes.
• Thompson sampling outperforms balance and weighted sampling in Table 2 (overall ACC 0.641 HCM, 0.574 CAM vs 0.533 and 0.540 for balance/weighted on CAM).

Surprising findings

• Full-batch RL degrades on rare classes: CAM ACC 0.100 and LVNC ACC 0.070 after FullBatchRL, showing majority bias persists under GRPO.
• Answering with disease names rather than multiple-choice options is reported to work better for training.

Limitations

• Test set is small (30 cases per class) and results averaged over three runs without reported variance.
• Localization bounding boxes are often oversized and localization training is limited to the MI subset.
• Evaluation relies partly on GPT-4o as judge with a small 50-case sample.

Experimental setup

Datasets: PMC-VQA CMR image-text pairs CMR-VQA (LGE-CMR cases)

Benchmarks: CMR-VQA test set (30 cases per class)

Metrics: ACC AUC F1

Baselines: GPT-4o Seed1.5-VL InternVL3-2B Qwen2.5-3B Lingshu-7B HuatuoGPT-Vision-7B MedGemma-4B Full-batch SFT

Builds on: GRPO PMC-VQA Thompson sampling Verl

Closest papers by meaning

• CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations (NeurIPS 2026 · similarity 0.91)
• Med-CMR: A Fine-Grained Benchmark Integrating Visual Evidence and Clinical Logic for Medical Complex Multimodal Reasoning (CVPR 2026 · similarity 0.88)
• M3CoTBench: Benchmark Chain-of-Thought of MLLMs in Medical Image Understanding (ICLR 2026 · similarity 0.87)
• MedReasoner: Reinforcement Learning Drives Reasoning Grounding from Clinical Thought to Pixel-Level Precision (AAAI 2026 · similarity 0.87)
• MedTVT-R1: A Multimodal LLM Empowering Medical Reasoning and Diagnosis (CVPR 2026 · similarity 0.87)
• Visual and Domain Knowledge for Professional-level Graph-of-Thought Medical Reasoning (ICML 2025 · similarity 0.86)