Impact of Defacing Procedures on Brain Age Gap Estimation
Poster No:
1151
Submission Type:
Abstract Submission
Authors:
Vivien Ivan1, Marius Vach1, Daniel Weiß1, Luisa Wolf1, Shammi More2, Kaustubh Patil3, Dennis Hedderich4, Nora Bittner5, Julian Caspers1, Christian Rubbert1
Institutions:
1Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital, Duesseldorf, Germany, 2Juelich Research Center, Juelich, Germany, 3Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich, Jülich, NRW, Germany, 4Technical University of Munich, Munich, Bavaria, Germany, 5Institute for Anatomy I, Medical Faculty, Heinrich-Heine University, Duesseldorf, Germany
First Author:
Vivien Ivan
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Co-Author(s):
Marius Vach
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Daniel Weiß
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Luisa Wolf
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Kaustubh Patil
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich
Jülich, NRW, Germany
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich
Jülich, NRW, Germany
Nora Bittner
Institute for Anatomy I, Medical Faculty, Heinrich-Heine University
Duesseldorf, Germany
Institute for Anatomy I, Medical Faculty, Heinrich-Heine University
Duesseldorf, Germany
Julian Caspers
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Christian Rubbert
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Department of Diagnostic and Interventional Radiology, Medical Faculty and University Hospital
Duesseldorf, Germany
Introduction:
Brain Age Gap Estimation (BrainAGE) is increasingly explored as an imaging biomarker for a variety of conditions, such as neurodegenerative diseases [Gaser 2013], or in the research of atypical aging, e.g., due to lifestyle risk factors [Bittner 2021]. BrainAGE is usually derived from applying machine-learning approaches to features derived from structural magnetic resonance images of the brain [Franke 2019; More 2023].
Removal of facial features from MRI scans is increasingly deemed mandatory from a data privacy perspective [Schwarz 2019]. However, defacing procedures may cause data integrity issues, e.g. affect regional brain volume measurements [Rubbert 2022], which could also compromise the quality of BrainAGE.
Thus, this study aims to systematically assess the impact of defacing on BrainAGE.
Removal of facial features from MRI scans is increasingly deemed mandatory from a data privacy perspective [Schwarz 2019]. However, defacing procedures may cause data integrity issues, e.g. affect regional brain volume measurements [Rubbert 2022], which could also compromise the quality of BrainAGE.
Thus, this study aims to systematically assess the impact of defacing on BrainAGE.
Methods:
A total of 364 Alzheimer's disease (AD) patients and 717 cognitively normal (CN) participants were included from the Alzheimer's Disease Neuroimaging Initiative (ADNI). For those, unaccelerated (AD n=290 (46.2% female, 75.13±7.79 years, 55-90 years), CN n=386 (49.5% female, 74.66±5.87 years, 56-89 years)), and accelerated 3D T1 imaging (AD n=203 (40.9% female, 74.76±8.02 years, 55-90 years), CN n=500 (61% female, 71.34±6.44 years, 55-90 years)) was available.
BrainAGE was derived using the best-performing model from [More 2023], trained on 2,953 non-ADNI healthy controls (18-88 years) from four large population-based studies. In essence, gray matter features were derived from standardized CAT12 for SPM12 preprocessing, smoothing with a 4 mm full-width-half-maximum kernel, resampling to 4 mm spatial resolution, and finally by principal component analysis for dimensionality reduction. Finally, Gaussian process regression was used to predict age, and BrainAGE was calculated by subtracting the chronological age from the predicted age.
BrainAGE was derived for each AD and CN before and after defacing using afni_refacer, fsl_deface, mri_deface, mri_reface, PyDeface, and spm_deface.
Furthermore, for AD n=74 (52.7% female, 75.51±8.19 years, 56-90 years) and CN n=84 (56% female, 75.93±4.97 years, 60-87 years) within-session unaccelerated repeat imaging was available, and BrainAGE differences between the two scans within the same session were calculated without any defacing to serve as a benchmark.
The difference in BrainAGE after defacing (BrainAGEdefaced – BrainAGEfullface), as well as the mean absolute error (MAE) and mean squared error (MSE) of BrainAGE with and without defacing were calculated separately for unaccelerated and accelerated imaging. Additionally, BrainAGE outliers due to defacing were identified using BrainAGE differences above the 75th or below the 25th percentile of the benchmark and Grubbs's test.
BrainAGE was derived using the best-performing model from [More 2023], trained on 2,953 non-ADNI healthy controls (18-88 years) from four large population-based studies. In essence, gray matter features were derived from standardized CAT12 for SPM12 preprocessing, smoothing with a 4 mm full-width-half-maximum kernel, resampling to 4 mm spatial resolution, and finally by principal component analysis for dimensionality reduction. Finally, Gaussian process regression was used to predict age, and BrainAGE was calculated by subtracting the chronological age from the predicted age.
BrainAGE was derived for each AD and CN before and after defacing using afni_refacer, fsl_deface, mri_deface, mri_reface, PyDeface, and spm_deface.
Furthermore, for AD n=74 (52.7% female, 75.51±8.19 years, 56-90 years) and CN n=84 (56% female, 75.93±4.97 years, 60-87 years) within-session unaccelerated repeat imaging was available, and BrainAGE differences between the two scans within the same session were calculated without any defacing to serve as a benchmark.
The difference in BrainAGE after defacing (BrainAGEdefaced – BrainAGEfullface), as well as the mean absolute error (MAE) and mean squared error (MSE) of BrainAGE with and without defacing were calculated separately for unaccelerated and accelerated imaging. Additionally, BrainAGE outliers due to defacing were identified using BrainAGE differences above the 75th or below the 25th percentile of the benchmark and Grubbs's test.
Results:
MAE and MSE for BrainAGE difference between the initial and repeat scan in the benchmark, both without defacing, were 1.15 and 2.25 for CN (BrainAGE difference -0.15±1.5), and 1.43 and 3.29 for AD (BrainAGE difference of 0.15±1.82).
Preprocessing with CAT12 for SPM12 or defacing failed in 1 afni_refacer and 275 mri_deface cases.
PyDeface performed best with an overall MAE of 0.33 and MSE of 0.27 (mean BrainAGE difference of 0.08±0.52, range -1.93-6.19, inter-quartile range -0.17–0.26 (see table in Figure 1) and the lowest number of outliers according to the benchmark (99, see table in Figure 1 and Figure 2). The overall MAE for fsl_deface and spm_deface was 0.35, for mri_reface 0.41, for mri_deface 0.63, and for afni_refacer 1.04.
The Grubbs's test identified 23 outliers after PyDeface, with spm_deface and mri_reface leading to a lower number of outliers (20 and 11, respectively, see table in Figure 1).
Preprocessing with CAT12 for SPM12 or defacing failed in 1 afni_refacer and 275 mri_deface cases.
PyDeface performed best with an overall MAE of 0.33 and MSE of 0.27 (mean BrainAGE difference of 0.08±0.52, range -1.93-6.19, inter-quartile range -0.17–0.26 (see table in Figure 1) and the lowest number of outliers according to the benchmark (99, see table in Figure 1 and Figure 2). The overall MAE for fsl_deface and spm_deface was 0.35, for mri_reface 0.41, for mri_deface 0.63, and for afni_refacer 1.04.
The Grubbs's test identified 23 outliers after PyDeface, with spm_deface and mri_reface leading to a lower number of outliers (20 and 11, respectively, see table in Figure 1).
·Figure 1
·Figure 2
Conclusions:
Brain Age Gap Estimation (BrainAGE) may be affected by defacing, however, in most approaches this influence is lesser than the variability observed in BrainAGE in repeat non-defaced imaging within the same session. Overall, PyDeface had negligible impact on BrainAGE, both in AD patients and healthy controls.
Lifespan Development:
Aging 1
Neuroinformatics and Data Sharing:
Databasing and Data Sharing
Informatics Other
Novel Imaging Acquisition Methods:
Anatomical MRI 2
Keywords:
Aging
Computational Neuroscience
Data analysis
MRI
Other - Data protection
1|2Indicates the priority used for review
Provide references using author date format
Franke K, Gaser C (2019): Ten Years of BrainAGE as a Neuroimaging Biomarker of Brain Aging: What Insights Have We Gained? Front Neurol 10:789.
Gaser C, Franke K, Klöppel S, Koutsouleris N, Sauer H, Initiative ADN (2013): BrainAGE in Mild Cognitive Impaired Patients: Predicting the Conversion to Alzheimer’s Disease. PLoS ONE 8:e67346.
More S, Antonopoulos G, Hoffstaedter F, Caspers J, Eickhoff SB, Patil KR, Initiative ADN (2023): Brain-age prediction: A systematic comparison of machine learning workflows. NeuroImage 270:119947.
Rubbert C, Wolf L, Turowski B, Hedderich DM, Gaser C, Dahnke R, Caspers J, Initiative ADN (2022): Impact of defacing on automated brain atrophy estimation. Insights Imaging 13:54.
Schwarz CG, Kremers WK, Therneau TM, Sharp RR, Gunter JL, Vemuri P, Arani A, Spychalla AJ, Kantarci K, Knopman DS, Petersen RC, Jr. CRJ (2019): Identification of Anonymous MRI Research Participants with Face-Recognition Software. N Engl J Med 381:1684–1686.