Bayesian Deep Learning Based Convolutional Neural Network for Classification of Parkinson’s Disease Using Functional Magnetic Resonance Images

Abstract

The primary aim of functional magnetic resonance imaging (fMRI) in a rest state is to analyze brain functioning using transient similarities in different areas of the brain as blood oxygenation level-dependent (BOLD) signals as a measure of synchronous neural activity. However, with the series of time-based data, numerous images of the patient’s brains provide sufficient knowledge to analyze their brain and encourage us to identify these disorders. We propose a new way of automatically extracting this data from non-time sequences using a Bayesian deep learning algorithm based on a convolutional neural network (CNN). We use whole data for the training phase instead of using those predefined points of interest (POIs)so that the trained model automatically ignores those brain points with no related information about the disease. Therefore, this method makes no assumptions regarding illness, patients, etc., and makes it possible to distinguish diseases that impact brain functioning by a universal disease diagnosis approach. This approach is a supervised algorithm that uses a small number of calculations using3D-CNN. Each fMRI scan (which involves t brain time slices) will be separated into t 3D images so that we can raise the amount of data set and simplify the calculations a lot. All these pictures were subsequently fed to a Bayesian network similar to LeNet-5 (but in 3D) to train our model. Then, to determine whether or not a person has Parkinson’s disease, we test his/her t fMRI images and get t different results which lead to a fraction of how unhealthful his/her brain is and if that fraction is above 0.5, we can classify that sample as a Parkinson’s patient.