Masked autoencoder pretraining for peripheral nerve segmentation in ultrasound images

초록

Despite recent advances in medical image segmentation, deep learning applications in musculoskeletal ultrasound remain limited by small labeled datasets and a narrow clinical focus. Most existing work centers on the median nerve, leaving other clinically relevant peripheral nerves understudied. This study explores masked autoencoder (MAE) pretraining as a scalable approach to overcome data scarcity in peripheral nerve segmentation. A dataset of 10,603 ultrasound images across eight nerves was constructed, spanning over 1,500 patients with normal nerve conduction studies. A SegFormer-based encoder was pretrained on unlabeled ultrasound images using MAE and subsequently fine-tuned on nerve-specific segmentation tasks. Pretraining was applied per cross-validation fold to prevent data leakage. Compared to randomly initialized baselines, MAE-pretrained models achieved consistent performance gains across all nerves. Improvements were most pronounced for low-resource targets such as the sural (0.144 → 0.399 Dice) and lateral femoral cutaneous (0.280 → 0.550 Dice) nerves. Nerves with more abundant training data, like the median (0.890 → 0.897 Dice) and ulnar (0.862 → 0.869 Dice), showed only marginal improvements. These findings indicate that MAE pretraining particularly strengthens feature extraction for small, difficult-to-label structures in ultrasound segmentation. This underscores the potential of self-supervised learning to ease annotation demands and support the development of more comprehensive clinical AI tools.

키워드