RELIEF: A structured multivariate approach for removal of latent inter-scanner effects
Poster No:
1965
Submission Type:
Abstract Submission
Authors:
Rongqian Zhang1, Lindsay Oliver2, Aristotle Voineskos2, Jun Young Park1
Institutions:
1University of Toronto, Toronto, Ontario, 2Centre for Addiction and Mental Health, Toronto, Ontario
First Author:
Co-Author(s):
Introduction:
Combining data collected from multiple study sites is becoming common and is advantageous in increasing the generalizability and replicability of scientific discoveries, but unwanted inter-scanner biases are commonly observed across neuroimaging data collected from multiple study sites or scanners. While several methods for handling such unwanted variations have been proposed, most of them use univariate approaches that could be too simple to capture all sources of scanner-specific variations. To address these challenges, we propose a novel multivariate harmonization method called RELIEF for estimating and removing both explicit and latent scanner effects. Our method is the first approach to introduce the simultaneous dimension reduction and factorization of interlinked matrices to a data harmonization context, which provides a new direction in methodological research for correcting inter-scanner biases.
Methods:
1. Approach: We assume that the data matrix consists of (i) covariate effects (ii) low-dimensional patterns independent to scanners, (iii) low-dimensional patterns specific to scanners, and (iv) noise with heterogeneous variances by scanners. We use a RELIEF to achieve such decomposition and harmonize data by dropping scanner-specific patterns and rescaling the noise.
2. Data: We analyzed fractional anistropy (FA) data collected by the Social Processes Initiative in Neurobiology of the Schizophrenia (SPINS) study, a large multi-site, multi-scanner study on examining social cognition in schizophrenia spectrum disorder. We extracted 73 FA values using the white matter atlas from the O'Donnell Research Group. For analysis, we grouped 172 subjects from the General Electric scanners (750w Discovery 3T or 750 Signa 3T) and179 subjects from the Siemens Prisma scanner.
3. Evaluations: We compared the performance of RELIEF to ComBat and CovBat, two existing batch-correction method in neuroimaging. Specifically, to evaluate covariance heterogeneity, we computed scanner-specific feature covariance matrices and compared their differences usi. We also used the Quadratic Discriminant Analysis (QDA) to predict scanners from the harmonized data. We obtained the averaged prediction accuracy using the leave-one-out cross validation (LOOCV).
2. Data: We analyzed fractional anistropy (FA) data collected by the Social Processes Initiative in Neurobiology of the Schizophrenia (SPINS) study, a large multi-site, multi-scanner study on examining social cognition in schizophrenia spectrum disorder. We extracted 73 FA values using the white matter atlas from the O'Donnell Research Group. For analysis, we grouped 172 subjects from the General Electric scanners (750w Discovery 3T or 750 Signa 3T) and179 subjects from the Siemens Prisma scanner.
3. Evaluations: We compared the performance of RELIEF to ComBat and CovBat, two existing batch-correction method in neuroimaging. Specifically, to evaluate covariance heterogeneity, we computed scanner-specific feature covariance matrices and compared their differences usi. We also used the Quadratic Discriminant Analysis (QDA) to predict scanners from the harmonized data. We obtained the averaged prediction accuracy using the leave-one-out cross validation (LOOCV).
·Conceptualization of the RELIEF method.
Results:
RELIEF identified non-ignorable latent batch effects from the Simens Prisma 3T scanner, and our subsequent analysis showed that it can be explained by additional site effects in Siemens Prisma. To visualize the efficiency of harmonization methods, we applied each method and computed the difference of sample covariance matrices between GE and Siemens. The Figure shows its superior performance to the others. We also used machine learning methods to evaluate how well RELIEF impaired the detection of scanners. RELIEF showed prediction accuracy (49.6% for FA, 61% for MD) to the random prediction, while results of ComBat (66.1% for FA, 83.2% for MD), CovBat (59.3% for FA, 82.6% for MD) implied remaining batch effects not being harmonized.
·Difference in sample covariances between Siemens Prisma 3T and General Eletric 3T after applying different harmonization methods to FA and MD. RELIEF shows superior performance in removing
Conclusions:
We developed a novel batch-correction method called RELIEF, that successfully captured latent patterns of batch effects specific to scanners. It is a multivariate approach based on dimension reduction that uses all features to capture heterogeneous covariance patterns specific to scanners. We showed superior performance of the RELIEF in removing scanner effects from the SPINS study and simulation studies. The proposed methodology is made publicly available as an R package (https://github.com/junjypark/RELIEF).
Modeling and Analysis Methods:
Diffusion MRI Modeling and Analysis
Exploratory Modeling and Artifact Removal 2
Methods Development
Multivariate Approaches 1
Novel Imaging Acquisition Methods:
Diffusion MRI
Keywords:
Computing
Data analysis
Machine Learning
Modeling
MRI
Multivariate
Open-Source Software
Statistical Methods
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review