Scale-attentional U-Net for the segmentation of the median nerve in ultrasound images

Abstract

Purpose: The aim of this study was to develop a neural network that accurately and effectively segments the median nerve in ultrasound (US) images. Methods: In total, 1,305 images of the median nerve of 123 normal subjects were used to train and evaluate the model. Four datasets from two measurement regions (wrist and forearm) of the nerve and two US machines were used. The neural network was designed for high accuracy by combining information at multiple scales, as well as for high efficiency to prevent overfitting. The model was designed in two parts (cascaded and factorized convolutions), followed by self-attention over scale and channel features. The precision, recall, dice similarity coefficient (DSC), and Hausdorff distance (HD) were used as performance metrics. The area under the receiver operating characteristic curve (AUC) was also assessed. Results: In the wrist datasets, the proposed network achieved 92.7% and 90.3% precision, 92.4% and 89.8% recall, DSCs of 92.3% and 89.7%, HDs of 5.158 and 4.966, and AUCs of 0.9755 and 0.9399 on two machines. In the forearm datasets, 79.3% and 87.8% precision, 76.0% and 85.0% recall, DSCs of 76.1% and 85.8%, HDs of 5.206 and 4.527, and AUCs of 0.8846 and 0.9150 were achieved. In all datasets, the model developed herein achieved better performance in terms of DSC than previous U-Net-based systems. Conclusion: The proposed neural network yields accurate segmentation results to assist clinicians in identifying the median nerve.