PROJECT TITLE :

Assessment of Parkinson’s Disease Severity from Videos using Deep Architectures

ABSTRACT:

The diagnosis of Parkinson's disease (PD) is made using clinical criteria, such as bradykinesia, rest tremor, rigidity, and other similar symptoms. The severity of Parkinson's disease symptoms can be evaluated using clinical rating scales; however, there is inter-rater variability in this process. In this article, we propose a deep learning-based automatic method for diagnosing Parkinson's disease (PD) that makes use of videos to aid in the clinical diagnosis process. We demonstrate the effectiveness of using a 3D convolutional neural network (CNN) as the baseline approach for the PD severity classification. We show that PD severity classification can benefit from transfer learning from non-medical datasets. This is because there is a lack of data in the clinical field, so we explore the possibility of transfer learning from non-medical datasets. We let the network focus more on the subtle temporal visual cues, such as the frequency of tremors, by designing a Temporal Self-Attention (TSA) mechanism. This was done so that we could bridge the domain discrepancy that exists between medical and non-medical datasets. The Movement Disorders Society's Unified Parkinson's Disease Rating Scale (MDS-UPDRS) part III is utilized in this study to investigate bradykinesia and postural tremors through the performance of seven tasks. In addition, we propose a multi-domain learning method as a means of determining the patient-level severity of PD through the utilization of task-assembling. Empirical evidence demonstrates that TSA and the task-assembling method are effective when applied to our PD video dataset. On the binary task level, we achieve the best MCC of 0.55, and on the three-class patient level, we achieve the best MCC of 0.39.