Towards Enhancing 3D Spatial Reasoning in Medical Multimodal Large Language Models

TL;DR — A data-centric method that boosts 3D volumetric medical understanding in multimodal LLMs by synthesizing radiologist-style chain-of-thought reasoning, letting a 2D-pretrained model rival costly native-3D architectures without 3D-specific pretraining.

• Introduces a large-scale structured reasoning dataset built via a "slice-wise data synthesis paradigm" that decomposes 3D reading into per-slice observations synthesized into an interpretable Chain-of-Thought, mirroring radiologists' diagnostic workflow.
• The synthesized reasoning enforces clinical principles: sequential spatial tracking, multi-slice spatial awareness for artifact mitigation, and differential exclusion.
• Instruction-tuning a standard 2D-pretrained MLLM on this data reportedly yields significant gains over the 2D baseline across multiple 3D medical benchmarks and rivals resource-intensive native 3D models.
• Positioned as avoiding expensive 3D pretraining/annotation; datasets and training workflows are released publicly (GitHub repo linked).