Using Neural Networks to Determine the Origin of Medical Ultrasound Images
MetadataShow full item record
This thesis investigates the possibility of using neural networks to determine the body location of medical ultrasound images. Neural networks were trained on several datasets of both synthetic and real ultrasound images, containing labels with the location of each image. Next, the networks predicted the location of unseen images, and the accuracy was measured. The first dataset consisted of images from three typical locations for ultrasound-guided regional anesthesia, where a classification accuracy of 85.3% was achieved. The second and third datasets consisted of synthetic ultrasound images, and neural networks were trained to predict coordinates for the images, using regression instead of classification. The networks predicted the coordinates with small errors, both in one and two dimensions. The fourth dataset consisted of real ultrasound recordings which covered a large area of the lower arm of nine test persons. Neural networks were trained to predict the 2-dimensional coordinates of the images by regression. In the best case, the average error of prediction was 1.66 𝑐𝑚. The final dataset consisted of images of the lower spine of three test persons. A neural network was trained to classify the images as sacrum, vertebra or intervertebral disc, and an average classification accuracy of 66.9% was achieved. The thesis shows that neural networks can be trained to predict locations of ultrasound images, both by classification and by regression, but more training data and further research is needed to increase the accuracy of the predictions. In addition, the thesis shows that pre-training on natural photographs can be applied to the ultrasound domain to help increase the accuracy of the neural networks.