Deep Learning-Based Segmentation of the Ulnar Nerve in Ultrasound Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate detection of the ulnar nerve (UN) in ultrasound (US) images is crucial for clinical diagnosis.
• Deep learning (DL) segmentation methods show promise for automated UN detection.
• A large-scale US dataset for UN segmentation is now available.

Purpose of the Study:

• To evaluate deep learning-based segmentation models for UN detection in US images.
• To compare the performance of various segmentation models and data augmentation techniques.
• To identify optimal strategies for UN segmentation using DL.

Main Methods:

• Utilized a large dataset of 4789 US images from 545 patients with expert-annotated UN segmentations.
• Compared six segmentation models with different backbone architectures.
• Analyzed the statistical significance of five data augmentation techniques (flipping, rotation, shearing, contrast/brightness, resizing).

Main Results:

• Shear, rotate, and resize augmentations significantly improved segmentation performance (p < 0.05).
• Traditional U-Net models achieved competitive results (Dice score: 0.88, IoU: 0.81).
• Newer architectures did not outperform traditional U-Net models for this task.

Conclusions:

• Systematic analysis provides insights into optimizing DL for UN segmentation.
• Data augmentation is critical for enhancing US-based nerve segmentation accuracy.
• U-Net remains a strong baseline for US nerve segmentation tasks.