Project Objectives
Detection of Brachial Plexus and Location
To allow users to detect the nerve structure called the brachial plexus on the ultrasound picture and the exact area in the patient’s body to implant the catheter.
The main challenge of the project was to train and set up the neural network to provide users with a high segmentation accuracy. This task involved a lot of time to spend on setting the neural network learning rate, its size and defining the optimal learning deviation (optimal loss function).
To solve this problem, we chose UNet architecture. It is a special neural network architecture developed for biomedical images segmentation. The hyper-parameters tuning has shown that the maximum accuracy in recognition is achieved through the combination of Dice + Binary Cross-entropy as a loss function.
As a result, a machine learning script connected with the neural network was developed. The ultrasound image of the shoulder joint comes as an input. As an output, there is an image with the highlighted brachial plexus. Since it is a script, it can be used in any type of app (web, desktop, mobile). The proper neural network training and setting has allowed getting ~70% accuracy in recognition.