Mayar Ali: Multi-scale convolutional neural networks for the detection of neurological and psychiatric diseases based on structural brain MRI data
BCCN Berlin / Charité-Universitätsmedizin Berlin
Recently deep learning neural networks have achieved state-of-the-art results in classifying neurological and psychiatric diseases using structural magnetic resonance imaging (sMRI) brain data. Most of the studies used either whole-brain or regions-based images for the disease classification. However, the impact of the different brain regions vary from one disease to another and some correlations between the brain regions remain unrevealed. In this study, we developed a new training strategy using hierarchical multi-scale convolutional neural networks (CNN) which combines both a region-of-interest (ROI) based approach and a whole-brain approach. This strategy acts as an attention guidance for better initialisation of the CNN focusing on the regions of high relevance for the disease of interest. We applied our proposed strategy on the public Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset and schizophrenia data provided by Prof. Schlagenhauf from Charité. We could get slightly better results than a randomly initialised baseline model with an increase of ∼1-4% in the accuracy. Our analysis also provided insights about potential regions of interest, such as hippocampus and fusiform gyrus for Alzheimer’s disease and 4th ventricle and frontal pole for schizophrenia. In addition, we could achieve relatively high accuracies using only a few brain regions instead of the whole brain image. We further extended our study to analyse the differences in the performance between female and male participants. Our results showed better classification accuracy among female participants with a difference of ∼10% compared to male participants. The suggested strategy can be further extended to different neurological and psychiatric diseases.
Master Thesis Defense
If you would like to join the online talk, please contact graduateprograms(at)bccn-berlin.de
Prof. Dr. Kerstin Ritter / Lisa Velenosi
Location: Zoom (Online)