Adapted 3D deep learning model for Parkinson's disease classification based on white matter changes

1397 

Abstract Submission 

Hyewon Shin¹, Yong Jeong¹

¹Korea Advanced Institute of Science & Technology, Daejeon, Republic of Korea

Hyewon Shin, BS.  
Korea Advanced Institute of Science & Technology
Daejeon, Republic of Korea

Yong Jeong, MD., PhD.  
Korea Advanced Institute of Science & Technology
Daejeon, Republic of Korea

Parkinson's disease (PD) is the world's second most common neurodegenerative disease, caused by dopaminergic neuronal loss in the substantia nigra due to the deposition of misfolded α-synuclein. As α-synuclein pathology has been suggested to cause early axonal changes that later result in neuronal degeneration, multiple previous studies have shown that diffusion tensor imaging (DTI), as an in vivo imaging technique, is capable of characterizing early detectable white matter (WM) changes in PD. However, another recent study suggested that DTI-based analyses may not be best suited for PD classification, as their binary support vector machine (bSVM) and multiple kernel learning (MKL) model demonstrated low accuracy. Recent advancements in the field of deep learning have given rise to models such as ResNet, designed specifically for image classification, with specialized convolutional layers and residual connections that handle grid-like data in a more efficient manner without the need for manual feature engineering. Therefore, we wondered whether we would be able to generate a DTI-based PD classification model demonstrating high classification accuracy if the machine learning technique itself was deeper and able to learn more complex baseline features. In this study, we generated a novel 3D deep learning classification model with a whole brain fractional anisotropy (FA) map input to classify Parkinson's disease patients from healthy controls with comparable accuracy to other recent state-of-the-art classification models.

625 healthy control (age, 61.4 ± 11.3; female, 37.3%) and 1402 cognitively unimpaired PD patient (age, 63.4 ± 9.4; female, 34.8%) data were selected from the PPMI database for training, validating and testing our model at a 7:2:1 ratio. The DTI and T1-weighted images were preprocessed using MRtrix3, FMRIB Software Library (FSL) and Advanced Normalization Tools (ANTs). The response function, estimating the signal expected in each voxel, was returned for each subject using the Dhollander algorithm. Using the returned response functions, we estimated the fiber orientation distributions (FODs) for each voxel, where the response function was used as a kernel in a contained spherical deconvolution operation. Using these FODs, we were able to estimate the fractional anisotropy (FA) values in each voxel. The returned 3D FA brain map acted as the input into the proposed 3D ResNet-34 PD classification model composed of novel skip connection structures in tandem with SE blocks. Pre-trained 2D ResNet-34 was converted into 3D by duplicating the 2D filters into the third dimension, and transfer learning was employed to reduce training time. Other layers were adjusted to match the 3D filters. Furthermore, the Taguchi method was employed to optimize the performance and improve the robustness of the model.

Our proposed model demonstrates a validation accuracy of 99.74% and a validation loss of 0.0418. Moreover, when evaluated on the test data set, the model demonstrated an accuracy of 95.35% and an AUC value of 0.98. Compared to most recent studies, which exhibit a range of accuracies between 49.4% and 98.2%, our model shows high accuracy, precision and recall. Furthermore, the model holds its own merits over other state-of-the-art models with higher classification accuracy in that it does not require multi-modal imaging data and manual feature extraction, which can easily increase computational cost and information loss.

We propose a novel 3D deep learning PD classification model with an accuracy that can compete against other state-of-the-art models. The main difference this model holds is its lack of need for manual feature extraction and multi-modal analysis to raise model accuracy. Furthermore, the proposed model has a strong advantage in that it can be applied to any other task with a 3D input, extending its value to characterizing other pathological changes and accommodating other imaging modalities.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ²

Classification and Predictive Modeling ¹

Diffusion MRI Modeling and Analysis

Aging

Degenerative Disease

Machine Learning

Movement Disorder

Open Data

White Matter

Other - Classification; Parkinson's Disease (PD); Diffusion Tensor Imaging (DTI); Deep Learning