Mapping White Matter Anomalies in Parkinson's Disease Progression using Deep Normative Tractometry

1605 

Abstract Submission 

Yixue Feng¹, Conor Owens-Walton¹, Bramsh Chandio¹, Julio Villalón-Reina¹, Corey McMillan², Peter Opriessnig³, Petra Schwingenschuh³, Duygu Tosun⁴, Sophia Thomopoulos¹, Ysbrand van der Werf⁵, Neda Jahanshad¹, Paul Thompson¹

¹University of Southern California, Los Angeles, CA, ²University of Pennsylvania, Philadelphia, PA, ³Medical University of Graz, Graz, Austria, ⁴University of California, San Francisco, San Francisco, CA, ⁵Vrije Universiteit Amsterdam, Amsterdam, Netherlands

Yixue Feng  
University of Southern California
Los Angeles, CA

Conor Owens-Walton, PhD  
University of Southern California
Los Angeles, CA

Bramsh Chandio  
University of Southern California
Los Angeles, CA

Julio Villalón-Reina  
University of Southern California
Los Angeles, CA

Corey McMillan, PhD  
University of Pennsylvania
Philadelphia, PA

Peter Opriessnig  
Medical University of Graz
Graz, Austria

Petra Schwingenschuh  
Medical University of Graz
Graz, Austria

Duygu Tosun  
University of California, San Francisco
San Francisco, CA

Sophia Thomopoulos  
University of Southern California
Los Angeles, CA

Ysbrand van der Werf  
Vrije Universiteit Amsterdam
Amsterdam, Netherlands

Neda Jahanshad, PhD  
University of Southern California
Los Angeles, CA

Paul Thompson, PhD  
University of Southern California
Los Angeles, CA

Normative models (NM) can be applied to brain metrics derived from structural, functional, and diffusion MRI to quantify individual deviations from the statistical distribution observed in a reference population, with applications in neurology and psychiatry. NMs have recently been extended to tractometry, using deep autoencoders (AE) to detect individual anomalies in white matter (WM) microstructure from diffusion MRI. While traditional NMs are typically estimated for single brain measures (e.g. hippocampal volume), AEs can encode statistical distributions of complex features such as 3D tract geometry as well as continuously varying measures along tracts. We have shown in our prior work that a Variational Autoencoder (VAE), the generative counterpart of AE, is able to capture structural features of tracts and can be used for anomaly detection in AD when trained on data from healthy controls (HC). In this work, we extend our framework to jointly model the 3D geometry of WM tracts and their microstructures, and show its ability to identify along-tract abnormalities across different stages of Parkinson's Disease (PD) in a multi-site cohort.

dMRI data from 112 HC and 437 participants with PD (age range: 31-87; 197 F/352 M) collected at 4 sites were preprocessed using the ENIGMA-DTI pipeline. Constrained spherical deconvolution reconstruction and the deterministic EuDX tracking method were used to generate whole-brain tractograms; 30 WM bundles were extracted using BUAN pipeline. To leverage a larger sample of HC, the VAE model was first pre-trained on 30 WM tracts from 198 HC of the TractoInferno dataset. Input features to the model contained 3D coordinates and fractional anisotropy (FA) mapped to 128 points per streamline in the native space. We then created 100 bootstrap samples from 50 HC participants for fine-tuning and calculated anomaly scores (mean absolute error, MAE) along 100 segments of each bundle for the remaining 62 HC and 437 PD subjects averaged across all the bootstrap sets (Fig. 1). ComBat harmonization was applied on along-tract MAE to adjust for site effects and linear regression was used to correct for age and sex effects. Group comparisons between HC and Hoehn-Yahr (HY) stage 1, 2, and 3 participants were conducted using a cluster-based permutation test on the t-statistic at alpha=0.05 to account for correlation between neighboring segments.

Since the VAE model was trained on HC participants, we expect to see larger deviations from the norm quantified using MAE scores in the PD participants. The cluster-based permutation test revealed significant group differences - in segments 89-100 in the right optic radiation (OR_R) between HC and HY2, segments 81-97 in the left uncinate fasciculus (UF_L), and segments 76-92 in the right inferior fronto-occipital fasciculus (IFOF_R) between HC and HY3 - with a medium effect calculated using Cohen's d (Fig. 2). While significant differences are detected in HY2 and HY3, the anomaly patterns emerge in earlier stages, most notable in IFOF_R at stage HY2.

Our normative tractometry framework based on VAE can encode both the WM macro- and micro-structures from HC participants. Given the limited sample size typical in this and other studies, our framework uses a public independent dataset for pretraining, and supports multi-site harmonization for along-tract statistics. In a multi-site PD cohort, our framework identified along-tract anomalies, supporting the notion that structural WM changes may be coupled with microstructural changes in PD. Future work will examine additional dMRI and shape metrics, along with the effect of the size, inclusion criteria and diversity of the reference population.

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

Diffusion MRI Modeling and Analysis ¹

Methods Development

Multivariate Approaches

Degenerative Disease

Machine Learning

Multivariate

Tractography

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC

Other - Deep Learning