On Friday the 12 June at 13.15 in MH:Hörmander Filip Winzell will defend his thesis "Machine Learning for Longitudinal Medical Data Analysis - Applications in Prostate Cancer and Alzheimer's Disease".

Abstract: 

The healthcare systems of today are facing large challenges, with increasing amounts of patients and overworked hospital staff. Prostate cancer and Alzheimer's disease are two of the most prevalent diseases, with incidence numbers expected to rise over the coming decade. Medical imaging plays a central role in current diagnostic procedures for these diseases, enabling the potential use of machine-learning-based image analysis. Overall, computer-aided diagnostics has seen a big increase in research over the last decade, with numerous applications where AI-based methods perform at the same level as experienced physicians. However, with the computational power available today, AI-based methods have the potential to achieve more in terms of prognostication and early detection of diseases. This is something that would be of high value for the treatment of the aforementioned diseases. Thus, the topic of this thesis is to investigate machine learning-based methods for the analysis of longitudinal medical imaging data, with the goal of improving the diagnostic procedures of prostate cancer and Alzheimer's disease.

Currently, there are no general screening programs for prostate cancer, despite the importance of early detection for successful treatment. Screening based on blood measures of prostate-specific antigen (PSA) has been shown to reduce mortality but also significantly increase the levels of over-treatment. To mitigate that, active surveillance has been suggested as an alternative to radical treatment following abnormal PSA values, where patients are monitored with recurring examinations. When a patient is deemed to have a high-risk prostate cancer, as defined by the Gleason grading scale, they receive treatment. However, the Gleason grading system is a subjective grading system with proven inter-observer variability. In this thesis, a method to predict the longitudinal treatment decision of prostate cancer patients on active surveillance is presented. It is based on the popular attention-based multiple instance learning framework, in combination with the state-of-the-art foundation model for pathology called UNI. This model achieved promising results, indicating that it is possible to reliably predict the onset of prostate cancer earlier than trained pathologists.  

Alzheimer's disease is a neurodegenerative disease, characterized by an abnormal accumulation of amyloid-$\beta$ and tau proteins in the brain. The cause of the disease is still unknown and the progression is highly heterogeneous across a population of patients. While it remains incurable, there are recently developed treatments that have been shown to effectively slow down the progression if the disease is detected early. Hence, to better understand why certain patients progress differently than others and how to detect them early is of high interest. In this thesis we developed an algorithm to find patterns of brain atrophy in Alzheimer's disease patients and connect them to other biological abnormalities. We found four distinct subtypes, that could explain parts of this highly complex disease.