Anatomical Footprint of the Impulse Control Disorders in Parkinson’s Disease: A Convolutional Vision Transformers Approach

Impulse Control Disorder in Parkinson’s patients are a psychiatric condition characterized by pathological gambling, compulsive buying, binge eating and hypersexuality. Even though that Impulse Control Disorders are attributed to dopaminergic treatment, recent studies associate this condition with abnormalities in brain anatomy. In this work, we used T1-weighted magnetic resonance images of three groups of Parkinson’s Disease subjects: 1) idiopathic Parkinson’s disease, 2) Pakinson’s patients with hypersexuality and eating impulsive control disorders, and 3) Parkinson’s patients presented compulsive buying and gambling. All the images were pre-processed including bias field correction, denoising, registration to the MNI template, and brain extraction. To identify each condition, transformer-based deep learning models were used. The Convolutional vision Transformer model uses 3D-convolutional layers to recognize patterns of Impulse Control Disorders in brain images. The 3D-convolutional layers were utilized to analyze the complete 3D brain anatomy and the larger receptive field as compared to 2D convolutions. Additionally, a ResNet model was implemented to compare performances. The ResNet model was based on connections between non-consecutive 3D convolutional layers. Testing results show that the transformer model was more efficient in terms of memory usage, training time, and classification accuracy compared to the ResNet model.