Performance benchmarking of deep learning models for real-time median nerve segmentation and cross-sectional area...

Median nerve, a major peripheral nerve, connects the hand to the central nervous system, facilitating upper limb motor function and sensation by transmitting sensory data from the palm and fingers. Damage to this nerve can result in motor and sensory deficits, with carpal tunnel syndrome (CTS) causing compression, leading to tingling and numbness in the thumb, index, middle, and lateral ring fingers.

This study aimed to develop an accurate deep-learning-based segmentation method for measuring the cross-sectional area (CSA) of the median nerve to facilitate the diagnosis of nerve entrapment syndromes and aid in surgical planning, with a focus on CTS.

This study introduces MNSeg-Net, a novel lightweight multiscale feature fusion network with 2.46M parameters for median nerve segmentation in ultrasound (US) frames, specifically designed to enable a fully automated, end-to-end clinical setup supporting real-time segmentation and CSA computation. The dataset comprised 100 subjects and 30 000 ultrasound frames, which were split into training (80%), validation (10%), and testing (10%) subsets with subject-wise separation to avoid data leakage. MNSeg-Net was benchmarked against state-of-the-art segmentation models, including UNet and its variants (UNet++ and U2Net). The performance was assessed using metrics such as the Dice similarity coefficient (DSC) and CSA difference. The statistical significance of performance differences was evaluated using paired t-tests, effect size (Cohen's d), and one-way ANOVA with Tukey's HSD correction for multiple comparisons at a $p$ -value threshold of 0.05, while statistical equivalence between models within predefined margins was formally assessed using the two one-sided test (TOST) procedure. Following quantitative validation, the model was deployed in a real-time clinical setup utilizing an Av.io HD Epiphan frame grabber to stream ultrasound images from the ultrasound machine to a GPU-equipped system. A secondary display running parallel to the original ultrasound screen visualized the segmented median nerve and computed the CSA values in real time.

MNSeg-Net achieved high segmentation performance, with average DSC scores of 94.7% at the wrist and 83.4% from the wrist to the elbow, and the lowest Hausdorff distance, matching the performance of the best-performing 44-million-parameter heavy U2Net model. Compared to U2Net, MNSeg-Net showed no statistically significant difference in DSC performance ( $p=9.11\times {10}^{-1}$ ; Cohen's $d=-0.003$ ; mean difference = -0.001), with formal equivalence testing confirming equivalence across all tested margins ( $\pm 0.01,\pm 0.03,\pm 0.05$ ). For CSA estimation, MNSeg-Net also showed no statistically significant difference from clinician-annotated values ( $p=1.14\times {10}^{-1}$ ; Cohen's $d=-0.041$ ; mean difference = -0.081), and equivalence was established at the $\pm 0.5$ margin, confirming a strong alignment with expert clinical assessments. MNSeg-Net demonstrated real-time performance by processing up to 43 frames per second on a single GPU, successfully segmenting the median nerve and computing CSA from ultrasound frames.

The developed MNSeg-Net-based clinical system represents an important step toward real-time median nerve assessment, enabling a fully automated solution for CTS diagnosis. By combining a lightweight architecture, real-time processing capability, and successful clinical deployment, it represents a substantial advancement in the CTS detection and management.