Efficient musculoskeletal annotation using free-form deformation.

Area of Science:

• Medical Imaging
• Deep Learning
• Anatomical Modeling

Background:

• Manual muscle segmentation for training deep learning models is labor-intensive and costly.
• Scalability of current methods is limited by the need for expert annotators.

Purpose of the Study:

• To develop and evaluate a novel tool for efficient, non-expert-assisted annotation of medical images for muscle segmentation.
• To assess the performance of deep learning models trained on datasets generated with this tool.

Main Methods:

• A user-friendly tool employing free-form deformation of a 3D anatomical template model was developed.
• Non-experts used the tool to fit the template to target magnetic resonance images (MRIs).
• An automatic segmentation network was trained using datasets created by the tool.

Main Results:

• Non-expert annotations achieved a Dice coefficient > 0.75 compared to expert segmentations.
• The tool-assisted segmentation showed minimal errors, such as mislabeling or omissions.
• Deep learning models trained with the tool-generated data performed comparably or better than those trained with expert data.

Conclusions:

• The developed tool significantly reduces time and labor costs for creating muscle segmentation datasets.
• This approach democratizes medical image annotation, enabling faster development of clinical deep learning applications.
• The tool facilitates efficient dataset creation for automatic muscle segmentation, potentially transforming medical image analysis.