Style transfer GANs for synthesizing controls for case-only studies in multi-study analyses
Poster No:
226
Submission Type:
Abstract Submission
Authors:
Shruti Gadewar1, Alyssa Zhu1, Iyad Ba Gari2, Sunanda Somu3, Sophia Thomopoulos4, Paul Thompson5, Talia Nir6, Neda Jahanshad7
Institutions:
1USC, Marina Del Rey, CA, 2University of Southern California, Marina Del Rey, CA, 3University of Southern California, Los Angeles, CA, 4USC, Marina del Rey, CA, 5Imaging Genetics Center, Keck School of Medicine of University of Southern California, Los Angeles, CA, 6University of Southern California Keck School of Medicine, Marina del Rey, CA, 7Imaging Genetics Center, Keck School of Medicine of University of Southern California, Los Angeles, California
First Author:
Co-Author(s):
Paul Thompson, PhD
Imaging Genetics Center, Keck School of Medicine of University of Southern California
Los Angeles, CA
Imaging Genetics Center, Keck School of Medicine of University of Southern California
Los Angeles, CA
Neda Jahanshad, PhD
Imaging Genetics Center, Keck School of Medicine of University of Southern California
Los Angeles, California
Imaging Genetics Center, Keck School of Medicine of University of Southern California
Los Angeles, California
Introduction:
Pooling data from multiple sites in neuroimaging studies enhances sample size, statistical power, and reproducibility [1]. Multi-study analyses require harmonization approaches to adjust for measurement variability in MRI acquisition protocols across studies. Statistical harmonization is often a first step after pooling subject level imaging features derived from raw scans across data collection sites [2]. This harmonization is often applied with respect to controls, and in studies that also include 'cases', the resulting harmonization parameters are then applied to the cases [3]. Some studies focus exclusively on variations within cases and lack controls, making harmonization challenging. To address this, we propose creating synthetic control T1-weighted (T1w) MRI for a target dataset by harmonizing the "style" of control images to that of target dataset 'cases'. We used ComBat-GAM [4] to statistically harmonize regional brain volumes across multiple Alzheimer's disease (AD) case/control datasets. We compared the effects of using either true or synthetic control T1w MRI brain volumes for statistical harmonization on resulting clinical associations.
Methods:
The AD datasets (Fig 1A) analyzed in this study were ADNI1, OASIS3, NACC and AIBL. 272 T1w MRI controls from OASIS1 and UK Biobank were age- and sex-matched to AIBL dementia cases (aged 60-81 years). We then created synthetic T1w control data for AIBL by harmonizing the "style" of OASIS1 and UKB images to AIBL dementia T1w. True AIBL control T1w images were used for validation. All T1w images were bias field corrected, skull-stripped using HD-BET [5], registered to MNI template using FSL's flirt [6] command with 9 degrees of freedom, and then zero-padded to 256x256x256 voxels. After "style" harmonization, images were moved back to subject space. Regional volumes for the thalamus, hippocampus, amygdala, putamen, caudate, accumbens, pallidum and ventricles were extracted from the T1ws for all datasets using FastSurfer [7]. Two ComBat-GAM harmonization models were trained with control volumes from ADNI1, OASIS3, NACC and either true or synthetic control volumes from AIBL. Site was used as the batch effect with age, sex and intracranial volume (ICV) covariates; age was specified as a nonlinear term. Linear mixed models were run to compare AIBL AD subcortical volumes to either 1) true AIBL, 2) synthetic AIBL, or 3) the original unharmonized OASIS and UKB control volumes. Linear regressions were also performed to compare control and AD subcortical volumes across ComBat harmonized studies; again, the use of AIBL true vs synthetic volumes was compared. Finally, we evaluated associations between ApoE4 count (0/1/2) and harmonized subcortical measures within the AD population; AD data was harmonized using either true or synthetic control volumes for comparison. All regressions included age, sex, ICV as fixed effect covariates.
Results:
Results are shown in Fig 2. In AIBL, AD cases had lower bilateral hippocampal volumes compared to both true and synthetic control participants (p=1.1✕10^-13 for true, p=2.6x10^-10 for synthetic). A paired t-test comparing pooled case-control effect sizes across all volumes when using true or synthetic AIBL control data was not significant (p=0.62) showing that both sets of regressions had similar results. Bilateral amygdala volume was found to be significantly associated with ApoE4 count in pooled AD participants, when true AIBL (r=-0.18; p=0.002) and synthetic AIBL control volumes (r=-0.19; p=0.003) were used for ComBat.
Conclusions:
This work is preliminary and has several limitations. We only evaluated synthetic controls in the case of AD, where the effect of neurodegeneration is more evident than in mood disorders or other psychiatric conditions. We performed a case-control analysis for validation, yet the primary objective of this work is to better allow for statistical harmonization of case-only datasets for case-only statistical analyses and avoid over-correction.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Modeling and Analysis Methods:
Other Methods 2
Keywords:
Degenerative Disease
Machine Learning
STRUCTURAL MRI
Other - Synthetic controls, ComBat, Harmonization, Generative adversarial network
1|2Indicates the priority used for review
Provide references using author date format
[2] J. Radua et al. (2020), “Increased power by harmonizing structural MRI site differences with the ComBat batch adjustment method in ENIGMA,” NeuroImage, vol. 218, p. 116956.
[3] R. Da-Ano et al. (2020), “Performance comparison of modified ComBat for harmonization of radiomic features for multicenter studies,” Sci. Rep., vol. 10, no. 1, p. 10248.
[4] R. Pomponio et al. (2020), “Harmonization of large MRI datasets for the analysis of brain imaging patterns throughout the lifespan,” NeuroImage, vol. 208, p. 116450.
[5] F. Isensee et al. (2019), “Automated brain extraction of multisequence MRI using artificial neural networks,” Hum. Brain Mapp., vol. 40, no. 17, pp. 4952–4964.
[6] M. Jenkinson et al. (2002), “Improved optimization for the robust and accurate linear registration and motion correction of brain images,” NeuroImage, vol. 17, no. 2, pp. 825–841.
[7] L. Henschel et al. (2020), “FastSurfer - A fast and accurate deep learning-based neuroimaging pipeline,” NeuroImage, vol. 219, p. 117012.
[8] M. Liu et al. (2023), “Style transfer generative adversarial networks to harmonize multisite MRI to a single reference image to avoid overcorrection,” Hum. Brain Mapp., vol. 44, no. 14, pp. 4875–4892.
Acknowledgements:
This work is supported in part by NIH grants: R01AG059874, RF1AG057892, P41EB015922, U01AG068057 and R01AG058854. This work was completed using UK Biobank Resource under application number 11559. Acknowledgments for OASIS and NACC can be found at (http://www.oasis-brains.org/#access, https://naccdata.org/publish-project/authors-checklist). Data used in the preparation of this article was obtained from the AIBL funded by the Commonwealth Scientific and Industrial Research Organization (CSIRO). AIBL researchers are listed at www.aibl.csiro.au. Data used in preparing this article were obtained from the ADNI database (adni.loni.usc.edu). As such, many investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete list of ADNI investigators: http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf.