MedVol-R1: Reward-Driven Evidence Grounding for Volumetric Reasoning Segmentation

MedVol-R1 introduces a reinforcement learning framework for volumetric reasoning segmentation in 3D medical imaging, decoupling evidence grounding from mask generation to improve interpretability and accuracy. The system uses an LVLM to identify key 2D evidence anchors before propagating them into coherent 3D segmentations, achieving state-of-the-art results on multiple medical imaging benchmarks without requiring expensive annotations.

MedVol-R1 addresses a critical limitation in medical image analysis by introducing architectural transparency to volumetric segmentation tasks. Traditional approaches collapse complex reasoning into opaque latent representations, making it difficult to audit clinical decision-making or generalize across varied clinical language. By explicitly grounding reasoning in verifiable 2D evidence before 3D reconstruction, the framework creates an auditable decision path that aligns with medical practice standards.

This work represents a significant advancement in medical AI interpretability. The decoupled architecture—where an LVLM identifies key axial slices and bounding boxes before a frozen MedSAM2 module performs volumetric delineation—mirrors how radiologists approach segmentation tasks. The use of reinforcement learning with multi-component reward functions eliminates the need for expensive chain-of-thought annotations while training the system to balance evidence informativeness, spatial accuracy, and volumetric coherence.

The clinical applications are substantial. Medical institutions increasingly require explainable AI systems for regulatory compliance and clinical adoption. MedVol-R1's transparent evidence anchoring provides the interpretability that hospital IT departments and clinicians demand, potentially accelerating deployment in real-world settings. The framework's strong performance on diverse datasets (CT-ORG, AbdomenCT-1K, KiTS23) suggests broad applicability across organ systems and imaging protocols.

Future development should focus on extending this approach to multimodal imaging scenarios and validating interpretability claims with radiologist studies. The framework's modular design suggests potential for fine-tuning domain-specific components while maintaining the interpretability benefits.

• →MedVol-R1 decouples evidence grounding from segmentation, creating interpretable decision pathways for clinical AI systems.
• →Reinforcement learning with multi-component rewards achieves state-of-the-art results without expensive chain-of-thought annotations.
• →The framework's verifiable 2D evidence anchors address regulatory and clinical adoption barriers for medical imaging AI.
• →Architecture design mirrors radiologist workflow, suggesting better generalization across diverse clinical narratives.
• →Performance improvements demonstrate clear gains of reinforcement learning over pure supervised fine-tuning approaches.