Anatomical Patterns Recognition of Impulse Control Disorders of Parkinsonian Patients Using Deep Learning of MRI Structural Images

In Parkinson’s disease (PD) several neuropsychiatric conditions are highly frequent including impulse control disorders (ICD). Driven by medication’s side-effects, these disorders lead to issues in self-control over rewarding cues that produce excessive behavior. Yet, objective means to detect and diagnose presence and severity of impulsivity in PD is lacking. The method developed in this work seeks to identify anatomical patterns in magnetic resonance images, which allow to classify PD with ICD vs PD patients using deep learning techniques. The proposed method includes pre-processing techniques to standardize magnetic resonance images. Different experiments using a variety of deep learning architectures were designed to classify the patterns. A GradCAM technique was also used for identifying the regions where the anatomical patterns were identified, which differentiate the two groups. We reveal an objective way to predict the presence of patterns associated to ICD in anatomical scans. Our model achieved an accuracy of 97.4% using Convolutional Neuronal Networks with an ensemble architecture; a novel objective clinical tool to diagnose ICD and better characterize detection of this severe neuropsychiatric problem.