A CNN-based classifier for probe to skin contact in a robotic ultrasound system

Abstract

Abstract:

In an ultrasound examination the sonographer must be precise and apply the right amount of pressure to get good quality images. An examination takes valuable time which could be reduced by introducing a robotic ultrasound system. In a robotic ultrasound system the ultrasound transducer needs to maintain good contact with the skin of the subject. Automatic skin to probe contact detection may therefore be useful in medical applications where it is desirable to use a robotic ultrasound device.

In this thesis a convolutional neural network (CNN) algorithm was implemented and used to distinguish if the ultrasound probe had full contact, partial contact or no contact with a person s skin. The main dataset consisted of ultrasound images of two body parts from one subject. Different subsets of the complete dataset was also used when training and testing the network. Four experiments with different subsets of the complete dataset were performed, with applied changes to the hyperparameters, in attempt to and the best overall CNN for the task.

The CNN was developed and trained using MATLAB R2018a with the Neural Network Toolbox. The results found indicated that a smaller CNN was sufficient for the task. The CNN which obtained the best results consisted of two convolutional layers and two fully connected layers. The experiment with this network was performed three times and reached an average accuracy of 100% and an average classification time of 0.0430 seconds. The accuracy achieved indicated that similarities in the training set and test set could have affected the reliability of the results. It is therefore important that in future work, it is acquired images from more body parts and more subjects to get enough variation in the dataset. This is important to be able to further improve the generalization of the CNN.