Abstract
Background and Objectives: We evaluate deep learning-based segmentation methods for detecting the ulnar nerve in ultrasound (US) images, leveraging the first-ever large US dataset of the ulnar nerve. We compare several widely used segmentation models, analyze their performance, and evaluate several common data augmentation techniques for the US. Materials and Methods: Our analysis is conducted on a large dataset of 4789 US images from 545 patients, with expert-annotated ground-truth segmentations of the ulnar nerve, and uses six segmentation models with several backbone architectures. Further, we analyze the statistical significance of five common data augmentation techniques on segmentation performance: flipping, rotation, shearing, contrast and brightness adjustments, and resizing. Results: In this study, the shear, rotate, and resize augmentations consistently improved segmentation performance across multiple runs, with p-values < 0.05 in a paired t-test relative to the no-augmentation baseline. Furthermore, we showed that newer architectures do not provide any metric improvements over traditional U-Net models, which achieved a Dice score of 0.88 and an IoU of 0.81. Conclusions: Through our systematic analysis of segmentation models and data augmentation strategies, we provide key insights into optimizing deep learning approaches for ulnar nerve segmentation and other US-based nerve segmentation tasks.