Identification of pelvic floor features in ultrasound images using an unsupervised machine learning algorithm

Finished: 2019-09-03

MSc assignment

Within the research of the master thesis, the goal is to analyze hidden features in ultrasound images using unsupervised autoencoder neural network. These networks are built to encode images into a lower dimension data, latent space, from which this latent space can be decoded into a reconstructed ultrasound image.

This latent space is a combination of features from the input image such that the output precisely can be reconstructed. This combination of features is such that a precise reconstruction is possible and thus all information is stored within this latent space. This latent space can be extracted from the neural network for the detection of any possible damage of the pelvic floor (avulsion). The detection of avulsion clusters can be done by currently known labels, but also similarly by the detection of new unique clusters.
These new clusters can give new insights of different aspects from the pelvic floor. The found clusters can be analyzed by looking at the ultrasound images and find the relation between the clustered images.