| dc.description.abstract | Schizophrenia (SCZ) is a complex mental disorder which affects about 1 in 300 people worldwide. Individuals with SCZ can experience psychosis in the form of hallucinations and delusions. The disorder has a severe impact on quality of life, not just for the patient but for their family and friends. Personalized treatment is rare and the selection of treatment often follows a "trial and error" regime where a self-report of symptoms decides what medication is most appropriate. Thus, there is a need to achieve a more personalized approach.
Since the discovery of functional magnetic resonance imaging (fMRI) in the 90s, functional neuroimaging has been used to investigate brain activity in patients with psychiatric disorders such as SCZ. Resting-state fMRI scans can be used to classify subjects with schizophrenia among a group of both patients and healthy controls using machine learning (ML). However, few have tried to classify the various subgroups of the disorder using ML. The aim of this thesis is to investigate if ML based on resting-state fMRI data (4D) can aid in subgrouping patients with SCZ. Furthermore, the feasibility of this approach to distinguish patients from healthy controls is investigated.
The approach consists of implementing a deep learning (DL) pipeline to handle four dimensional data, before analysing both online (N=148) and local (N=316) data. The study also assesses how different preprocessing of images impact DL models and explores hyperparameter combinations to optimize the performance of models.
The problem addressed in the thesis is difficult and probably an ill-posed problem with more variability than what would be ideal for sub-grouping in SCZ. Therefore, the overall performance of the implemented ML models are lower than expected. However, to our knowledge, this is the first attempt on distinguishing between data based on 4D neuroimaging data directly. The project shows that minimizing preprocessing, as well as using data only from one source rather than grouping several datasets, is beneficial. Hyperparameter selection improves performance and could potentially be further optimized and explored to improve performance. The proposed approach is able to reproduce previous attempts of separation between patients and controls, even if the analysis is performed on the raw data directly (4D) rather than feature extracted fMRI data. Thus, the proposed approach could still be valuable in clinical research and clinical follow up in the future. | |
