The recent advancement of ultrasound imaging has helped clinicians obtain high resolution images of peripheral nerves using high and ultra-high frequency probes. This has allowed clinicians to study peripheral nerve infrastructures and assess different pathological states. However, without proper training and experience in using ultrasound, it can be challenging in identifying the nerves on ultrasound images. One solution is to develop an automated ultrasound image analysis platform using deep learning, to enhance physicians’ ability to identify nerve structures. Convolutional Neural Networks (CNN), a subset of deep learning, can achieve this goal as it has been widely used in performing image segmentation and object detection.

We obtained ultrasound videoclips of the median nerve at the forearm and the wrist from 4 healthy volunteers. A batch of 500 frames of the forearm, and a separate batch of 500 frames of the wrist were extracted from the videoclips. The median nerve in each frame was manually traced using Fiji, an open-source platform. We divided each batch of frames randomly into a ‘training’ and ‘testing’ ratio of 4:1. A CNN model called U-Net was trained on the training dataset. The model performance on the testing dataset was evaluated using mean Intersection-over-Union (IoU) and Dice scores.

For the forearm, we obtained a mean IoU score of 0.855 (standard deviation SD: 0.058) and a Dice score of 0.921 (SD: 0.036). For the wrist, we obtained a mean IoU score of 0.907 (SD: 0.011) and a Dice score of 0.951 (SD: 0.006).

Our model showed that deep learning can identify the median nerve automatically with high accuracy. We will perform a similar analysis of the ulnar, fibular and tibial nerves and in larger populations.