SegmentAnyBone: A universal model that segments any bone at any location on MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Magnetic Resonance Imaging (MRI) provides high-quality, non-invasive insights into the human body.
• Accurate segmentation of tissues in MRI is crucial for precise diagnosis and treatment planning.
• Bone segmentation in MRI is challenging, with limited publicly available algorithms and a focus on specific anatomical areas.

Purpose of the Study:

• To develop a versatile, publicly available deep learning model for bone segmentation in MRI across multiple standard locations.
• To enable more accurate quantitative assessments of musculoskeletal conditions through improved bone segmentation.
• To provide a model that operates in both fully automated and prompt-based segmentation modes.

Main Methods:

• Collected and annotated a novel MRI dataset (320 volumes, >10k slices) across diverse protocols and anatomical regions.
• Investigated various standard network architectures and strategies for automated segmentation.
• Developed SegmentAnyBone, a foundation model extending the Segment Anything Model (SAM), for versatile bone segmentation.

Main Results:

• The proposed SegmentAnyBone model demonstrated effective bone segmentation across different anatomical locations and MRI sequences.
• Comparative analysis showed competitive or superior performance against previous approaches.
• Generalization analysis confirmed the model's robustness on external datasets.

Conclusions:

• SegmentAnyBone offers a versatile and publicly available solution for bone segmentation in MRI.
• The model enhances quantitative musculoskeletal analysis, addressing a gap in current radiological practice.
• The foundation model approach provides a flexible and powerful tool for medical image segmentation.