3DReasonKnee: Avance del Razonamiento Fundamentado en Modelos de Visión y Lenguaje Médicos

Área de la Ciencia:

• Medical Imaging AI
• Computer Vision
• Clinical Decision Support

Sus antecedentes:

• Current vision-language models (VLMs) lack the ability to ground anatomical regions in 3D medical images and perform step-by-step reasoning, hindering clinical adoption.
• Existing 3D datasets do not support the ; grounded reasoning; required for realistic diagnostic workflows and trustworthy clinician-AI collaboration.

Objetivo del estudio:

• To introduce 3DReasonKnee, the first dataset enabling 3D grounded reasoning for medical images.
• To facilitate the development of VLMs capable of localized, step-by-step diagnostic assessment in 3D medical volumes.
• To establish a benchmark for evaluating VLM performance in anatomical localization and diagnostic reasoning.

Principales métodos:

• Developed 3DReasonKnee, a dataset comprising 7,970 3D knee MRI volumes and 494,000 quintuples.
• Each quintuple includes MRI volume, diagnostic question, 3D bounding box, clinician-generated reasoning steps, and severity assessment.
• Created ReasonKnee-Bench for evaluating VLM localization and diagnostic accuracy, benchmarking five state-of-the-art VLMs.

Principales resultados:

• Established a novel benchmark (ReasonKnee-Bench) for evaluating 3D grounded reasoning in medical VLMs.
• Provided baseline performance metrics for five leading VLMs on localization and diagnostic accuracy tasks.
• Demonstrated the potential for improved VLM performance in clinically relevant 3D medical image analysis.

Conclusiones:

• 3DReasonKnee is a unique resource for advancing multimodal medical AI, capturing orthopedic surgeons' diagnostic expertise.
• The dataset and benchmark facilitate the development of AI systems capable of 3D, clinically aligned, localized decision-making.
• Future work can leverage 3DReasonKnee to enhance clinician-AI collaboration and diagnostic trust.