Unsupervised lesion detection with multi-view MRI and autoencoders

Prostate Cancer (PCa) is one of the most common health threats in the developed world and the most diagnosed cancer among men in Norway. Detecting a lesion is a challenging task, and early diagnosis and treatment can be crucial for the patient. Computer-assisted systems using Deep Learning (DL) to analyze Magnetic Reso- nance Imaging (MRI) data could help improve diagnostic accuracy if used together with current practices. However, the DL systems require large amounts of annotated data which are rarely available. In addition, the presence of factors such as bias in the form of a higher prevalence of one class in classification tasks can harm the final results of the developed system. This thesis explores methods for performing anomaly detection on MRI image data of the prostate, where lesions are the anomalies. We tackle the lack of annotated data by using unsupervised learning and a type of neural network called Autoencoders (AEs). Specifically, Convolutional Autoencoders (CAEs), Variational Autoencoders (VAEs), and multi-view CAEs are compared. We evaluate the methods by means of Area Under the Curve (AUC) to measure the ability to separate the images with a lesion from those that do not. The results of this thesis show that using a multi-view CAE offers advantages over the single-view based CAE and VAE in terms of ROC-AUC and that a fully unsupervised approach holds potential for lesion detection in prostate MRI.