Slice-Level Reasoning Aims to Strengthen 3D Medical MLLMs

Introduction

Multimodal large language models have made visible progress on 2D medical images, but 3D volumetric imaging remains a harder problem. In CT or MRI reading, clinicians do not simply inspect one frame; they track findings across adjacent slices, separate real anatomical changes from artifacts, and weigh alternative explanations. The arXiv paper “Towards Enhancing 3D Spatial Reasoning in Medical Multimodal Large Language Models” focuses on this missing capability: how to give a 2D-pretrained medical MLLM a stronger sense of 3D spatial reasoning without relying on expensive 3D-specific pretraining.

Key ideas

• Moving beyond rigid answers: Many medical vision-language datasets are built as fixed VQA pairs or final reports. These formats may contain the answer, but they often hide the reasoning path. For volumetric interpretation, that missing intermediate logic matters.

• Slice-wise data synthesis: The authors introduce a structured reasoning dataset built through a slice-level synthesis paradigm. Instead of treating a scan as an opaque volume, the method translates global clinical priors into fine-grained observations for individual slices, then integrates them into an interpretable Chain-of-Thought.

• Closer to radiology workflow: The synthesized reasoning is designed to enforce clinical principles such as sequential spatial tracking, multi-slice awareness for reducing artifact-driven mistakes, and differential exclusion. In other words, the model is trained to follow a process more similar to how radiologists move through a 3D study.

• Data-centric improvement of 2D models: Rather than building a large native 3D model from scratch, the paper instruction-tunes a standard 2D-pretrained MLLM using the synthesized data. Evaluations across multiple 3D medical benchmarks show significant gains over the 2D baseline, with the resulting model approaching more resource-intensive native 3D architectures.

Why it matters

The contribution is important because 3D medical AI faces a practical data problem. High-quality volumetric annotations are expensive, and clinical datasets are often difficult to share or inspect. If slice-level synthesis can make implicit diagnostic logic explicit, it may offer a more scalable way to train models for volumetric understanding.

The work should not be read as proof that such systems are ready for autonomous diagnosis. The reliability of synthesized reasoning, generalization across diseases and scanners, and the clinical faithfulness of generated reasoning chains all need careful validation. Still, the direction is notable: medical MLLMs are moving from recognizing isolated images toward modeling spatial processes across volumes. If the released repository, datasets, and workflows are reproducible, this line of work could become a useful bridge between 2D vision-language models and practical 3D medical imaging applications.

Source: arXiv