Improved cerebrovascular brain-age accuracy by multi-sequence, multi-center harmonisation
Poster No:
2601
Submission Type:
Abstract Submission
Authors:
Mathijs Dijsselhof1,2, Candace Moore3, Wibeke Nordhøy4, Dani Beck5,6,7, Lars Westlye5,6,8, Nishi Chaturvedi9,10, Alun Hughes9, David Cash11,12, Jonathan Schott11, Frederik Barkhof1,2,13, James Cole14,15, Henk Mutsaerts1,2, Jan Petr1,16
Institutions:
1Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Vrije Universiteit, Amsterdam, Netherlands, 2Amsterdam Neuroscience, Brain Imaging, Amsterdam, Netherlands, 3Netherlands eScience Center, Amsterdam, Netherlands, 4Physics and Computational Radiology, Radiology and Nuclear Medicine, Oslo University Hospital, Oslo, Norway, 5Norwegian Centre for Mental Disorders Research (NORMENT), Oslo University Hospital, Oslo, Norway, 6Department of Psychology, University of Oslo, Oslo, Norway, 7Division of Mental Health and Substance Abuse , Diakonhjemmet Hospital, Oslo, Norway, 8KG Jebsen Centre for Neurodevelopmental Disorders, University of Oslo, Oslo, Norway, 9MRC Unit for Lifelong Health & Ageing, Department of Population Science, University College London, London, United Kingdom, 10Population Sciences and Experimental Medicine, Institute of Cardiovascular Science, University College London, London, United Kingdom, 11Dementia Research Centre, UCL Queen Square Institute of Neurology, London, United Kingdom, 12UK Dementia Research Institute at UCL, University College London, London, United Kingdom, 13Queen Square Institute of Neurology and Centre for Medical Image Computing, University College London, Amsterdam, Netherlands, 14Dementia Research Centre, Queen Square Institute of Neurology, UCL, London, United Kingdom, 15Centre for Medical Imaging Computing, Computer Science, UCL, London, United Kingdom, 16Helmholtz-Zentrum Dresden-Rossendorf, Institute of Radiopharmaceutical Cancer Research, Dresden, Germany
First Author:
Mathijs Dijsselhof
Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Vrije Universiteit|Amsterdam Neuroscience, Brain Imaging
Amsterdam, Netherlands|Amsterdam, Netherlands
Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Vrije Universiteit|Amsterdam Neuroscience, Brain Imaging
Amsterdam, Netherlands|Amsterdam, Netherlands
Co-Author(s):
Wibeke Nordhøy
Physics and Computational Radiology, Radiology and Nuclear Medicine, Oslo University Hospital
Oslo, Norway
Physics and Computational Radiology, Radiology and Nuclear Medicine, Oslo University Hospital
Oslo, Norway
Dani Beck, Dr
Norwegian Centre for Mental Disorders Research (NORMENT), Oslo University Hospital|Department of Psychology, University of Oslo|Division of Mental Health and Substance Abuse , Diakonhjemmet Hospital
Oslo, Norway|Oslo, Norway|Oslo, Norway
Norwegian Centre for Mental Disorders Research (NORMENT), Oslo University Hospital|Department of Psychology, University of Oslo|Division of Mental Health and Substance Abuse , Diakonhjemmet Hospital
Oslo, Norway|Oslo, Norway|Oslo, Norway
Lars Westlye
Norwegian Centre for Mental Disorders Research (NORMENT), Oslo University Hospital|Department of Psychology, University of Oslo|KG Jebsen Centre for Neurodevelopmental Disorders, University of Oslo
Oslo, Norway|Oslo, Norway|Oslo, Norway
Norwegian Centre for Mental Disorders Research (NORMENT), Oslo University Hospital|Department of Psychology, University of Oslo|KG Jebsen Centre for Neurodevelopmental Disorders, University of Oslo
Oslo, Norway|Oslo, Norway|Oslo, Norway
Nishi Chaturvedi
MRC Unit for Lifelong Health & Ageing, Department of Population Science, University College London|Population Sciences and Experimental Medicine, Institute of Cardiovascular Science, University College London
London, United Kingdom|London, United Kingdom
MRC Unit for Lifelong Health & Ageing, Department of Population Science, University College London|Population Sciences and Experimental Medicine, Institute of Cardiovascular Science, University College London
London, United Kingdom|London, United Kingdom
Alun Hughes
MRC Unit for Lifelong Health & Ageing, Department of Population Science, University College London
London, United Kingdom
MRC Unit for Lifelong Health & Ageing, Department of Population Science, University College London
London, United Kingdom
David Cash
Dementia Research Centre, UCL Queen Square Institute of Neurology|UK Dementia Research Institute at UCL, University College London
London, United Kingdom|London, United Kingdom
Dementia Research Centre, UCL Queen Square Institute of Neurology|UK Dementia Research Institute at UCL, University College London
London, United Kingdom|London, United Kingdom
Jonathan Schott
Dementia Research Centre, UCL Queen Square Institute of Neurology
London, United Kingdom
Dementia Research Centre, UCL Queen Square Institute of Neurology
London, United Kingdom
Frederik Barkhof, MD, Ph. D
Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Vrije Universiteit|Amsterdam Neuroscience, Brain Imaging|Queen Square Institute of Neurology and Centre for Medical Image Computing, University College London
Amsterdam, Netherlands|Amsterdam, Netherlands|Amsterdam, Netherlands
Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Vrije Universiteit|Amsterdam Neuroscience, Brain Imaging|Queen Square Institute of Neurology and Centre for Medical Image Computing, University College London
Amsterdam, Netherlands|Amsterdam, Netherlands|Amsterdam, Netherlands
James Cole, PhD
Dementia Research Centre, Queen Square Institute of Neurology, UCL|Centre for Medical Imaging Computing, Computer Science, UCL
London, United Kingdom|London, United Kingdom
Dementia Research Centre, Queen Square Institute of Neurology, UCL|Centre for Medical Imaging Computing, Computer Science, UCL
London, United Kingdom|London, United Kingdom
Henk Mutsaerts
Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Vrije Universiteit|Amsterdam Neuroscience, Brain Imaging
Amsterdam, Netherlands|Amsterdam, Netherlands
Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Vrije Universiteit|Amsterdam Neuroscience, Brain Imaging
Amsterdam, Netherlands|Amsterdam, Netherlands
Jan Petr
Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Vrije Universiteit|Helmholtz-Zentrum Dresden-Rossendorf, Institute of Radiopharmaceutical Cancer Research
Amsterdam, Netherlands|Dresden, Germany
Radiology and Nuclear Medicine, Amsterdam University Medical Centers, Vrije Universiteit|Helmholtz-Zentrum Dresden-Rossendorf, Institute of Radiopharmaceutical Cancer Research
Amsterdam, Netherlands|Dresden, Germany
Introduction:
Accelerated brain ageing is associated with increased risks of dementia and mortality, and can be assessed with machine learning and structural brain MRI data to determine the predicted age difference (brain-PAD) between biological and chronological age[1]. Advanced brain-PAD approaches aim to increase accuracy and sensitivity to specific pathologies by adding sequences able to assess cerebrovascular health, a contributor to cognitive decline[2].
Improved brain-PAD accuracy, and classification of Alzheimer's Disease (AD) patients, were shown with Cerebrovascular Brain-age by adding Arterial Spin Labeling (ASL) perfusion MRI to T1w and FLAIR data in a single-site study[3,4]. Such a pre-trained model has, however, a limited value for data from different MRI vendors or acquired with different parameters. While combining (ASL) studies is often a prerequisite to reach sufficient sample sizes for machine learning, acquisition differences can lead to a systematic bias of brain-PAD. The impact of commonly used harmonisation techniques on this bias is not sufficiently studied.
Here, we assessed the influence of feature harmonisation with NeuroCombat[5] on Cerebrovascular brain-PAD, trained in a large dataset with T1w, FLAIR, and ASL sequences and applied in clinical datasets differing in sequence parameters and age range.
Improved brain-PAD accuracy, and classification of Alzheimer's Disease (AD) patients, were shown with Cerebrovascular Brain-age by adding Arterial Spin Labeling (ASL) perfusion MRI to T1w and FLAIR data in a single-site study[3,4]. Such a pre-trained model has, however, a limited value for data from different MRI vendors or acquired with different parameters. While combining (ASL) studies is often a prerequisite to reach sufficient sample sizes for machine learning, acquisition differences can lead to a systematic bias of brain-PAD. The impact of commonly used harmonisation techniques on this bias is not sufficiently studied.
Here, we assessed the influence of feature harmonisation with NeuroCombat[5] on Cerebrovascular brain-PAD, trained in a large dataset with T1w, FLAIR, and ASL sequences and applied in clinical datasets differing in sequence parameters and age range.
Methods:
Imaging data MRI data was obtained from three population-based studies: NORMENT[4] (GE, n=1107, 54.9% female, 49.7±16.9 years, 18-95 years) with PCASL 3D spiral-FSE (PLD=2025ms, LD=1450ms), SABRE[6] (Philips, n=727, 53% male, 71.2±6.6 years, 37-90 years) with PCASL 2D EPI (PLD=2000ms, LD=1800ms), and Insight46[7] (Siemens, n=282, 50.7% female, 70.6±0.66 years, 69–72 years) with PCASL 3D GraSE (PLD=1800ms, LD=1800ms). Additionally, structural (T1w and FLAIR) images were included. Data was pre-processed with ExploreASL (v1.10).
Machine learning Brain-age model features were created from GM and WM volumetrics (T1w), (log-transform) WMH volumetrics (FLAIR), and vascular territory-based CBF and spatial coefficient of variation (sCoV; ASL)3. Training was performed using the ExtraTrees regressor (scikit-learn 1.3.2) on NORMENT, validated using 5-fold cross-validation, and tested on Insight46 and SABRE. Dataset features were harmonised using NeuroCombat[5], and performance between unharmonised (UH) and harmonised (H) features was assessed by comparing brain-PAD and the mean (MAE) and standard deviation of the absolute brain-PAD.
Machine learning Brain-age model features were created from GM and WM volumetrics (T1w), (log-transform) WMH volumetrics (FLAIR), and vascular territory-based CBF and spatial coefficient of variation (sCoV; ASL)3. Training was performed using the ExtraTrees regressor (scikit-learn 1.3.2) on NORMENT, validated using 5-fold cross-validation, and tested on Insight46 and SABRE. Dataset features were harmonised using NeuroCombat[5], and performance between unharmonised (UH) and harmonised (H) features was assessed by comparing brain-PAD and the mean (MAE) and standard deviation of the absolute brain-PAD.
Results:
Harmonisation decreased mean differences between all groups in all individual features (Figure 1). The harmonisation had no effect on the results (UH/H MAE = 4.8±3.9/4.7±3.8 years; p=0.87) within the validation dataset. Without harmonisation, brain-PADs were -14.0 ± 5.1 (Insight46) and -10.6 ± 6.6 (SABRE) years. Harmonisation reduced brain-PAD (p<0.001) in both datasets to -2.6±4.7 and -4.1±5.9 years, respectively, and also reduced MAE significantly to 3.9±3.6 (Insight46) and 5.7±4.4 (SABRE) years.
·Figure 1: Imaging features of T1w, FLAIR and ASL sequences, per dataset, before and after harmonisation.
·Figure 2: Boxplots showing brain-PAD in years (row 1), and absolute brain-PAD in years (row 2) for every (un)harmonised dataset.
Conclusions:
Harmonisation effects between Insight46 and SABRE affected the ASL features the most, perhaps due to more variation in the different ASL sequence parameters having a direct impact on CBF, compared to the structural scans[8]. The remaining differences, especially in GM volume and CBF, could be attributed to dataset age differences. Harmonisation improved Brain-PAD and MAE values for Insight46 and SABRE, with more similar results to NORMENT, with brain-PADS of Insight46 agreeing with previous literature (-2.8±8.0 years)[9]. Both testing datasets showed negative average brain-PADs, perhaps due to known brain-age prediction biases at an older age[10].
To conclude, harmonisation reduces sequence- and site-specific biases in brain-age predictions using structural and, especially, ASL features. Further efforts should investigate applications in broader age ranges with a range of cognitive function to ensure that variance due to pathology and heterogeneous ageing remains present after harmonisation.
To conclude, harmonisation reduces sequence- and site-specific biases in brain-age predictions using structural and, especially, ASL features. Further efforts should investigate applications in broader age ranges with a range of cognitive function to ensure that variance due to pathology and heterogeneous ageing remains present after harmonisation.
Lifespan Development:
Aging 2
Modeling and Analysis Methods:
Methods Development
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Anatomy and Functional Systems
Novel Imaging Acquisition Methods:
Imaging Methods Other
Physiology, Metabolism and Neurotransmission :
Cerebral Metabolism and Hemodynamics 1
Keywords:
Aging
Cerebral Blood Flow
Cerebrovascular Disease
Cognition
Data analysis
Degenerative Disease
Machine Learning
MRI
STRUCTURAL MRI
Other - Arterial Spin Labelling MRI
1|2Indicates the priority used for review
Provide references using author date format
[2] Iadecola, C. et al., (2019) Neurovascular and Cognitive Dysfunction in Hypertension. Circ. Res. 124, 1025–1044.
[3] Dijsselhof, M. B. J. et al. (2023) The value of arterial spin labelling perfusion MRI in brain age prediction. Hum. Brain Mapp. 44, 2754–2766.
[4] Rokicki, J. et al. (2021) Multimodal imaging improves brain age prediction and reveals distinct abnormalities in patients with psychiatric and neurological disorders. Hum. Brain Mapp. 42, 1714–1726.
[5] Fortin, J.-P. et al. (2017) Harmonization of multi-site diffusion tensor imaging data. Neuroimage 161, 149–170.
[6] Jones, S. et al. (2020) Cohort Profile Update: Southall and Brent Revisited (SABRE) study: a UK population-based comparison of cardiovascular disease and diabetes in people of European, South Asian and African Caribbean heritage. Int. J. Epidemiol. 49, 1441–1442e.
[7] James, S.-N. et al. (2018) Using a birth cohort to study brain health and preclinical dementia: recruitment and participation rates in Insight 46. BMC Res. Notes 11, 885.
[8] Baas, K. P. A. et al. (2021) Effects of Acquisition Parameter Modifications and Field Strength on the Reproducibility of Brain Perfusion Measurements Using Arterial Spin-Labeling. AJNR Am. J. Neuroradiol. 42, 109–115.
[9] Wagen, A. Z. et al. (2022) Life course, genetic, and neuropathological associations with brain age in the 1946 British Birth Cohort: a population-based study. Lancet Healthy Longev 3, e607–e616.
[10] de Lange, A.-M. G. et al., (2020) Commentary: Correction procedures in brain-age prediction. NeuroImage. Clinical vol. 26 102229.