Read like a radiologist: Efficient vision-language model for 3D medical imaging interpretation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Medical vision-language models (VLMs) show promise in 2D image interpretation but struggle with 3D medical imaging due to computational demands and data limitations.
• Existing 3D VLMs often use sub-volumetric features, leading to correlated representations and neglecting crucial slice-specific details in 3D medical images.

Purpose of the Study:

• To introduce MS-VLM, a novel model designed to overcome the limitations of current 3D medical vision-language models.
• To develop a VLM that mimics the human radiologist's workflow for 3D medical image interpretation, capturing inter-slice dependencies effectively.

Main Methods:

• MS-VLM utilizes self-supervised 2D transformer encoders to learn volumetric representations from sequences of slice-specific features.
• The model processes 3D medical images without sub-volumetric patchification, allowing flexibility with varying slice lengths and multiple imaging planes/phases.

Main Results:

• MS-VLM demonstrated superior performance in radiology report generation on chest CT and rectal MRI datasets.
• The model produced more coherent and clinically relevant reports compared to existing methods.
• MS-VLM effectively captures inter-slice dependencies, improving volumetric representation from 3D medical images.

Conclusions:

• MS-VLM represents a significant advancement in 3D medical image interpretation.
• The model's ability to mimic radiologists' workflows enhances the robustness and clinical relevance of medical VLMs for 3D imaging.