Evaluating Deep Learning-Based Nerve Segmentation in Brachial Plexus Ultrasound Under Realistic Data Constraints

Accurate nerve localization is critical for the success of ultrasound-guided regional anesthesia, yet manual identification remains challenging due to low image contrast, speckle noise, and inter-patient anatomical variability. This study evaluates deep learning-based nerve segmentation in ultrasound images of the brachial plexus using a U-Net architecture, with a focus on how dataset composition and annotation strategy influence segmentation performance. We find that training on combined data from multiple ultrasound machines (SIEMENS ACUSON NX3 Elite and Philips EPIQ5) provides regularization benefits for lower-performing acquisition sources, though it does not surpass single-source training when matched to the target domain. Extending the task from binary nerve segmentation to multi-class supervision (artery, vein, nerve, muscle) results in decreased nerve-specific Dice scores, with performance drops ranging from 9% to 61% depending on dataset, likely due to class imbalance and boundary ambiguity. Additionally, we observe a moderate positive correlation between nerve size and segmentation accuracy (Pearson r=0.587, p<0.001), indicating that smaller nerves remain a primary challenge. These findings provide methodological guidance for developing robust ultrasound nerve segmentation systems under realistic clinical data constraints.

💡 Research Summary

This paper investigates the practical deployment of deep‑learning‑based segmentation of the brachial plexus in ultrasound‑guided regional anesthesia, focusing on realistic data constraints. Using a public dataset of 1,052 grayscale ultrasound frames from 101 patients, the authors separate images into two groups that correspond to two commercial ultrasound machines (SIEMENS ACUSON NX3 Elite and Philips EPIQ5) based on visual characteristics and quantitative image metrics (brightness, contrast, tonal richness, and especially sharpness). After patient‑level splitting, standard preprocessing (cropping, resizing, grayscale conversion) and modest data augmentation (±10° rotation, up to 12.5 % translation, 0.85–1.25 scaling) are applied.

A vanilla U‑Net is chosen as a stable baseline; binary cross‑entropy loss is used for single‑class (nerve vs background) training, while categorical cross‑entropy is employed for four‑class training (nerve, artery, vein, muscle). Training follows a five‑fold GroupK‑Fold scheme with Adam (lr = 1e‑3), early stopping (patience = 10), ReduceLROnPlateau, and a batch size of 32, ensuring that any performance differences stem from data composition or annotation strategy rather than architectural tweaks.

Three research questions are addressed. First, combining data from both machines yields a regularization benefit for the lower‑quality device: the Dice score for the Philips‑derived subset improves from 0.58 (single‑source) to 0.64 when mixed with Siemens data. However, when the target domain is matched (training and testing on the same machine), single‑source training still outperforms mixed training, indicating that domain‑specific data remain optimal for maximal performance.

Second, extending supervision from binary to multi‑class segmentation degrades nerve‑specific Dice scores by 9 %–61 % across experiments. The authors attribute this drop to severe class imbalance (nerve pixels constitute <5 % of the image) and ambiguous boundaries between nerve, vessels, and surrounding muscle. While multi‑class labels provide richer anatomical context, they compromise the precision of the nerve mask unless additional techniques such as class‑weighting, focal loss, or oversampling are introduced.

Third, a moderate positive correlation (Pearson r = 0.587, p < 0.001) is observed between measured nerve diameter and segmentation accuracy, confirming that smaller nerves (≈2 mm or less) are substantially harder to segment. This finding underscores the need for higher‑resolution acquisition, super‑resolution post‑processing, or nerve‑specific prior models to improve performance on diminutive structures.

The study’s limitations include the manual, visual assignment of images to device groups due to missing metadata, the lack of diverse patient factors (e.g., obesity, probe pressure) that affect real‑world image quality, and the exclusive focus on U‑Net without benchmarking lightweight or domain‑adaptation architectures.

In summary, the work demonstrates that (1) heterogeneous data can act as a regularizer for weaker devices but does not replace the advantage of device‑matched training, (2) multi‑class supervision trades off nerve‑specific accuracy for broader anatomical awareness, and (3) nerve size is a key determinant of segmentation success. These insights provide concrete guidance for developers of clinical AI tools who must balance limited annotation resources, device heterogeneity, and the need for reliable nerve localization in ultrasound‑guided anesthesia.