Related Experiment Video
Updated: Jun 11, 2026
06:26
Direct Linear Transformation for the Measurement of In-Situ Peripheral Nerve Strain During Stretching
Published on: January 12, 2024
Masked autoencoder pretraining for peripheral nerve segmentation in ultrasound images
Matthew Webster1,2, Ko Eun Kim3, Jaewon Lim4
1Chung-Ang University Industry-Academic Cooperation Foundation, Seoul, South Korea.
Scientific Reports
|June 9, 2026
Summary
Masked autoencoder (MAE) pretraining significantly improves peripheral nerve segmentation in ultrasound images, especially for understudied nerves with limited data. This self-supervised approach enhances AI tool development for clinical applications.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurosurgery
Background:
- Deep learning in musculoskeletal ultrasound segmentation faces challenges due to limited labeled data and a narrow clinical focus.
- Current research predominantly studies the median nerve, neglecting other critical peripheral nerves.
Purpose of the Study:
- To investigate the efficacy of masked autoencoder (MAE) pretraining for enhancing peripheral nerve segmentation in ultrasound images.
- To address data scarcity and improve segmentation performance for understudied peripheral nerves.
Main Methods:
- A dataset of 10,603 ultrasound images from 1,500+ patients across eight peripheral nerves was curated.
- A SegFormer encoder underwent MAE pretraining on unlabeled ultrasound data, followed by fine-tuning for nerve segmentation.
- Pretraining was implemented per cross-validation fold to prevent data leakage.
Main Results:
- MAE pretraining consistently improved segmentation performance across all evaluated nerves compared to random initialization.
- Significant performance gains were observed for low-resource nerves like the sural (Dice 0.144 to 0.399) and lateral femoral cutaneous (Dice 0.280 to 0.550).
- Nerves with larger datasets, such as the median and ulnar, showed only minor improvements.
Conclusions:
- MAE pretraining effectively enhances feature extraction for segmenting small and challenging structures in ultrasound images.
- Self-supervised learning, particularly MAE, shows promise in reducing annotation burdens and advancing clinical AI tools for comprehensive nerve assessment.
