-
Synthesizing study-specific controls using generative models on open access datasets for harmonized multi-study analyses
Authors:
Shruti P. Gadewar,
Alyssa H. Zhu,
Iyad Ba Gari,
Sunanda Somu,
Sophia I. Thomopoulos,
Paul M. Thompson,
Talia M. Nir,
Neda Jahanshad
Abstract:
Neuroimaging consortia can enhance reliability and generalizability of findings by pooling data across studies to achieve larger sample sizes. To adjust for site and MRI protocol effects, imaging datasets are often harmonized based on healthy controls. When data from a control group were not collected, statistical harmonization options are limited as patient characteristics and acquisition-related…
▽ More
Neuroimaging consortia can enhance reliability and generalizability of findings by pooling data across studies to achieve larger sample sizes. To adjust for site and MRI protocol effects, imaging datasets are often harmonized based on healthy controls. When data from a control group were not collected, statistical harmonization options are limited as patient characteristics and acquisition-related variables may be confounded. Here, in a multi-study neuroimaging analysis of Alzheimer's patients and controls, we tested whether it is possible to generate synthetic control MRIs. For one case-control study, we used a generative adversarial model for style-based harmonization to generate site-specific controls. Downstream feature extraction, statistical harmonization and group-level multi-study case-control and case-only analyses were performed twice, using either true or synthetic controls. All effect sizes using synthetic controls overlapped with those based on true study controls. This line of work may facilitate wider inclusion of case-only studies in multi-study consortia.
△ Less
Submitted 29 February, 2024;
originally announced March 2024.
-
A Comprehensive Corpus Callosum Segmentation Tool for Detecting Callosal Abnormalities and Genetic Associations from Multi Contrast MRIs
Authors:
Shruti P. Gadewar,
Elnaz Nourollahimoghadam,
Ravi R. Bhatt,
Abhinaav Ramesh,
Shayan Javid,
Iyad Ba Gari,
Alyssa H. Zhu,
Sophia Thomopoulos,
Paul M. Thompson,
Neda Jahanshad
Abstract:
Structural alterations of the midsagittal corpus callosum (midCC) have been associated with a wide range of brain disorders. The midCC is visible on most MRI contrasts and in many acquisitions with a limited field-of-view. Here, we present an automated tool for segmenting and assessing the shape of the midCC from T1w, T2w, and FLAIR images. We train a UNet on images from multiple public datasets t…
▽ More
Structural alterations of the midsagittal corpus callosum (midCC) have been associated with a wide range of brain disorders. The midCC is visible on most MRI contrasts and in many acquisitions with a limited field-of-view. Here, we present an automated tool for segmenting and assessing the shape of the midCC from T1w, T2w, and FLAIR images. We train a UNet on images from multiple public datasets to obtain midCC segmentations. A quality control algorithm is also built-in, trained on the midCC shape features. We calculate intraclass correlations (ICC) and average Dice scores in a test-retest dataset to assess segmentation reliability. We test our segmentation on poor quality and partial brain scans. We highlight the biological significance of our extracted features using data from over 40,000 individuals from the UK Biobank; we classify clinically defined shape abnormalities and perform genetic analyses.
△ Less
Submitted 1 May, 2023;
originally announced May 2023.
-
3D Grid-Attention Networks for Interpretable Age and Alzheimer's Disease Prediction from Structural MRI
Authors:
Pradeep Lam,
Alyssa H. Zhu,
Iyad Ba Gari,
Neda Jahanshad,
Paul M. Thompson
Abstract:
We propose an interpretable 3D Grid-Attention deep neural network that can accurately predict a person's age and whether they have Alzheimer's disease (AD) from a structural brain MRI scan. Building on a 3D convolutional neural network, we added two attention modules at different layers of abstraction, so that features learned are spatially related to the global features for the task. The attentio…
▽ More
We propose an interpretable 3D Grid-Attention deep neural network that can accurately predict a person's age and whether they have Alzheimer's disease (AD) from a structural brain MRI scan. Building on a 3D convolutional neural network, we added two attention modules at different layers of abstraction, so that features learned are spatially related to the global features for the task. The attention layers allow the network to focus on brain regions relevant to the task, while masking out irrelevant or noisy regions. In evaluations based on 4,561 3-Tesla T1-weighted MRI scans from 4 phases of the Alzheimer's Disease Neuroimaging Initiative (ADNI), salience maps for age and AD prediction partially overlapped, but lower-level features overlapped more than higher-level features. The brain age prediction network also distinguished AD and healthy control groups better than another state-of-the-art method. The resulting visual analyses can distinguish interpretable feature patterns that are important for predicting clinical diagnosis. Future work is needed to test performance across scanners and populations.
△ Less
Submitted 18 November, 2020;
originally announced November 2020.
-
A Fast, Accurate Two-Step Linear Mixed Model for Genetic Analysis Applied to Repeat MRI Measurements
Authors:
Qifan Yang,
Gennady V. Roshchupkin,
Wiro J. Niessen,
Sarah E. Medland,
Alyssa H. Zhu,
Paul M. Thompson,
Neda Jahanshad
Abstract:
Large-scale biobanks are being collected around the world in efforts to better understand human health and risk factors for disease. They often survey hundreds of thousands of individuals, combining questionnaires with clinical, genetic, demographic, and imaging assessments; some of this data may be collected longitudinally. Genetic associations analysis of such datasets requires methods to proper…
▽ More
Large-scale biobanks are being collected around the world in efforts to better understand human health and risk factors for disease. They often survey hundreds of thousands of individuals, combining questionnaires with clinical, genetic, demographic, and imaging assessments; some of this data may be collected longitudinally. Genetic associations analysis of such datasets requires methods to properly handle relatedness, population structure and other types of biases introduced by confounders. Most popular and accurate approaches rely on linear mixed model (LMM) algorithms, which are iterative and computational complexity of each iteration scales by the square of the sample size, slowing the pace of discoveries (up to several days for single trait analysis), and, furthermore, limiting the use of repeat phenotypic measurements. Here, we describe our new, non-iterative, much faster and accurate Two-Step Linear Mixed Model (Two-Step LMM) approach, that has a computational complexity that scales linearly with sample size. We show that the first step retains accurate estimates of the heritability (the proportion of the trait variance explained by additive genetic factors), even when increasingly complex genetic relationships between individuals are modeled. Second step provides a faster framework to obtain the effect sizes of covariates in regression model. We applied Two-Step LMM to real data from the UK Biobank, which recently released genoty** information and processed MRI data from 9,725 individuals. We used the left and right hippocampus volume (HV) as repeated measures, and observed increased and more accurate heritability estimation, consistent with simulations.
△ Less
Submitted 15 March, 2019; v1 submitted 29 October, 2017;
originally announced October 2017.
