Image Prior Transfer and Ensemble Architectures for Parkinson's Disease Detection

EasyChair Preprint 6676

Abstract

Neural networks have shown promising results in many applications including computer aided diagnosis systems. However, insufficient effort has been expended on model knowledge transfer combined with ensemble architecture structures. Here, our use case focuses on detecting Parkinson's Disease (PD) by automatic pattern recognition in brain magnetic resonance (MR) images. In order to train a robust neural network, sufficiently large amount of labeled MR image data is essential. However, this is challenging because ground truth data needs to be labeled by clinical experts, who often have busy daily schedules. Furthermore, brain MR images are not often captured for PD patients. Therefore, we explore the effectiveness of pre-training neural networks using natural images instead of brain MR images of PD patients. We also propose different ensemble architecture structures, and demonstrate that they outperform existing models on PD detection. Experimental results show that our detection performance is significantly better compared to models without prior training using natural images. This finding suggests a promising direction when no or insufficient MR image training data is available. Furthermore, we performed occlusion analysis to identify the brain regions that the models focused on to deliver higher performance on PD detection during the decision making process.

Keyphrases: Magnetic Resonance Imaging, Magnetic Resonance Imaging., Model Knowledge Transfer, Parkinson's Disease Detection, deep learning, ensemble learning

@booklet{EasyChair:6676,
  author    = {Tahjid Ashfaque Mostafa and Irene Cheng},
  title     = {Image Prior Transfer and Ensemble Architectures for Parkinson's Disease Detection},
  howpublished = {EasyChair Preprint 6676},
  year      = {EasyChair, 2021}}