Assessment of Deep Learning Classification Methods for Parkinsonism

SUMMARY The growth rate in the number of people diagnosed with Parkinsonism is substantial. Estimates indicate that from 1990 to 2015 the number of Parkinsonism diagnoses doubled, with more than 6 million people currently diagnosed. By 2040, there will be between 12-14 million people diagnosed with Parkinsonism. Parkinson’s disease (PD), multiple system atrophy Parkinsonian variant (MSAp), and progressive supranuclear palsy (PSP) are neurodegenerative forms of Parkinsonism, which can be difficult to diagnose as they share similar motor and non-motor features, and they each have an increased chance of developing dementia. In the first five years of a PD diagnosis, about 58% of PD are misdiagnosed, and of these misdiagnoses about half have either MSA or PSP. Since PD, MSAp, and PSP require unique treatment plans and different medications, and clinical trials testing new medications require the correct diagnosis, there is an urgent need for clinic ready diagnostic level markers for differential diagnosis of PD, MSAp, and PSP. A promising approach to identify different forms of Parkinsonism is diffusion magnetic resonance imaging (dMRI), as there is no contrast drug, the technique is safe and is already used clinically in traumatic brain injury and stroke. The data collection takes 6-12 minutes and is compatible on current 3 Tesla MRI systems worldwide. Based on academic research at University of Florida, Automated Imaging Diagnostics, LLC is developing a commercial software package using free-water diffusion imaging as an innovative biomarker to help in the diagnosis of PD, MSAp, and PSP. The software currently distinguishes PD, MSAp, and PSP with over 90% accuracy, and can achieve this accuracy on different scanner manufactures. Our next goal in this Phase I project is to further improve the innovation and accuracy of our software technology by employing deep learning classification algorithms for the diagnosis of Parkinsonism. The specific aim of this current Phase I project is to substitute and compare the use of our existing Support Vector Machine (SVM) method with two different Residual Deep Neural Network (ResDNN) architectures for estimating disease type (PD/MSAp/PSP) through the following two milestones. First, in Milestone 1 we will determine if a ResDNN method that processes the same feature vector as our SVM solution improves the accuracy for differentiating a) PD and atypical Parkinsonism (MSAp/PSP) and b) MSAp and PSP by 5%. Second, in Milestone 2, we will determine if a ResDNN method that processes directly the raw input image data (instead of our derived feature vector) improves the accuracy for differentiating a) PD and atypical Parkinsonism (MSAp/PSP) and b) MSAp and PSP by 5%. This Phase I project will facilitate our long-term objective of developing a high-precision diagnostic software that can be used by radiologists for diagnosing different types of Parkinsonism.