Supervisors: Prof. Dr. Ender Konukoglu, Prof. Dr. Luc Van Gool, PD Dr. med. Olivio F. Donati
Statistically, one out of six men develops Prostate Cancer (PCa) in their lifetime. Early detection and accurate diagnosis facilitate the adequate treatment and can therefore improve the survival rate. Much research has been conducted to apply latest machine learning insights on multi-parametric Magnetic Resonance Images (mp-MRI) over the past years. While we could observe a trend from classical machine learning techniques (e.g. SVM) to end-to-end learning (i.e. mainly Neural Networks), the learning objective has not changed. Common approaches focus on a Region of Interest (ROI) learning objective, i.e. to classify parts of the prostate. On the contrary, whole gland analysis aims to classify the state of the whole prostate gland. Whole gland analysis yields several benefits, such as easier data collection for supervised learning. This Master Thesis aims to analyze prostate glands as a whole based on the Prostate Imaging Reporting and Data System (PI-RADS) v2 score, which aggregates the PCa severity of a prostate into a single number between 1 and 5. For this purpose, a dataset consisting of 3D mp-MRIs of 609 prostates was collected and preprocessed for the use in a Neural Network. Architectures using 3D convolution were tested as a baseline. Furthermore, a more complex architecture based on 2D convolutions, which enables the use of 2D pretrained networks, was developed. Neither a 2D implementation trained from scratch, nor fine-tuning or freezing a pretrained network could outperform the 3D approach with an AUC of 0.68 for PI-RADS ≤ 3 or > 3. This work marks a first step towards PI-RADS v2 score prediction and underlines difficulties, such as the substantial variation in the PI-RADS v2 assessment.