Variability of brain-age estimates in MS within and between three different MR scanners
Poster No:
1448
Submission Type:
Abstract Submission
Authors:
Lonneke Bos1, David Van Nederpelt1, James Cole2, Hugo Vrenken1, Eva Strijbis3, Bastiaan Moraal1, Joost Kuijer1, Bernard Uitdehaag1, Alle Meije Wink1, Frederik Barkhof1, Bas Jasperse1
Institutions:
1Amsterdam University Medical Centre, Amsterdam, Netherlands, 2University College London, London, London, 3Department of Neurology, MS Center Amsterdam, Amsterdam Neuroscience, Amsterdam UMC, Amsterdam, Netherlands
First Author:
Co-Author(s):
Eva Strijbis
Department of Neurology, MS Center Amsterdam, Amsterdam Neuroscience, Amsterdam UMC
Amsterdam, Netherlands
Department of Neurology, MS Center Amsterdam, Amsterdam Neuroscience, Amsterdam UMC
Amsterdam, Netherlands
Introduction:
Neurodegenerative processes due to normal aging change the macroscopic structure of the brain, which are similar to brain changes observed in neurodegenerative brain diseases. "Accelerated" aging of the brain is therefore regarded as a marker of disease-related neurodegeneration. Recent deep and machine learning techniques trained on aging healthy subjects estimate the subjects age from standard brain MRI scans. By subtracting the person's true age from this estimated brain-age the brain predicted age difference (brain-PAD) is acquired. Brain-PAD is related to clinical and radiological evidence of disease severity in multiple sclerosis (MS). Previous research showed that people with MS (pwMS) have a brain-PAD of 7.3 ± 0.8 (SD) years, on average. The intra- and inter-scanner reliability of brain-PAD estimation is currently unknown and is important for interpreting brain-PAD and for clinical implementation. The aim of this project is to investigate within-scanner repeatability and between-scanner reproducibility of brain-PAD using three different brain-age models and same-day scan-rescan imaging on three different scanners.
Methods:
For 30 pwMS and 10 age and sex-matched HC (mean age 44.2 ± 11.7 years and 39.2 ± 12.9 years, respectively), MS with EDSS range: 0.0–6.5, 3D T1-weighted MRI scans were acquired on a GE Discovery MR750 (3T), a Siemens Sola (1.5T) and a Siemens Vida (3T) scanner (AMS2 dataset). Each person was scanned twice on all scanners in one day. Brain-PAD was determined using the MIDI model (DenseNet121), the brainageR model (Gaussian Process Regression) and DeepBrainNet (inception-resnetv2). Repeatability and reproducibility were assessed using intraclass correlation coefficient (ICC, ICC absolute agreement within-scanner and ICC consistency between-scanners). Another way to measure repeatability and reproducibility was the smallest detectable change (SDC).
Results:
Within-scanner repeatability was excellent (ICC>0.93) for all three models for both HC and pwMS (see Fig 1.A). Within-scanner repeatability was better for HC (SDC range 1.56 - 3.06 years) than for pwMS (SDC range 2.63 – 4.08 years) (see Table 1.A) for all three models.
Between-scanner ICC for brainageR and the MIDI model was good to excellent (>0.85), for both HC and pwMS. For DeepBrainNet the ICC between-scanners was between 0.15 and 0.32 when the GE was compared with the Sola or the Vida, but was >0.92 for comparison between the Sola and the Vida (see Fig 1.B). Between-scanner SDC for the MIDI model was 5.06 years for HC and 5.73 years for pwMS. Between-scanner SDC for brainageR between 6.50 years for HC and 6.86 years for pwMS. Between-scanner consistency was lower for the brainageR model for GE vs Sola, than for GE vs Vida. For DeepBrainNet the SDC between-scanners was 24.27 years for HC and 22.08 years for pwMS (see Table 1.B).
Between-scanner consistency was lower for the brainageR model when comparing the Sola (1.5T) with the 3T scanners. This might be due to the lower signal to noise ratio as a result of the lower field strength.
Differences in scanner characteristics and imaging protocols cause an imbalance of within-scanner repeatability and between-scanner reproducibility. Model fine-tuning could minimize discrepancies without sacrificing one for the other. The SDC within-scanner is clinically acceptable, whereas the SDC between-scanners is too large to detect the increase of brain-age for pwMS. This makes it difficult to use brain-age prediction on an individual level to predict disease progression.
Between-scanner ICC for brainageR and the MIDI model was good to excellent (>0.85), for both HC and pwMS. For DeepBrainNet the ICC between-scanners was between 0.15 and 0.32 when the GE was compared with the Sola or the Vida, but was >0.92 for comparison between the Sola and the Vida (see Fig 1.B). Between-scanner SDC for the MIDI model was 5.06 years for HC and 5.73 years for pwMS. Between-scanner SDC for brainageR between 6.50 years for HC and 6.86 years for pwMS. Between-scanner consistency was lower for the brainageR model for GE vs Sola, than for GE vs Vida. For DeepBrainNet the SDC between-scanners was 24.27 years for HC and 22.08 years for pwMS (see Table 1.B).
Between-scanner consistency was lower for the brainageR model when comparing the Sola (1.5T) with the 3T scanners. This might be due to the lower signal to noise ratio as a result of the lower field strength.
Differences in scanner characteristics and imaging protocols cause an imbalance of within-scanner repeatability and between-scanner reproducibility. Model fine-tuning could minimize discrepancies without sacrificing one for the other. The SDC within-scanner is clinically acceptable, whereas the SDC between-scanners is too large to detect the increase of brain-age for pwMS. This makes it difficult to use brain-age prediction on an individual level to predict disease progression.
·ICC for three brain-age methods for both HC and pwMS, for within-scanner repeatability and between-scanner reproducibility
·Smallest Detectable Change (SDC) for three brain-age methods for both HC and pwMS, within- and between-scanners
Conclusions:
Within-scanner reproducibility was excellent for all brain-PAD models for both HC and pwMS. The brainageR model was most robust between-scanners, while DeepBrainNet was most robust within-scanners. The MIDI model showed overall the best results for repeatability and reproducibility. While within-scanner repeatability showed promising results, caution is warranted in determining brain-age across different scanners.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)
Lifespan Development:
Aging 2
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Other Methods
Keywords:
Aging
Demyelinating
MRI
Neurological
STRUCTURAL MRI
Other - Brain age; multiple sclerosis; artificial intelligence
1|2Indicates the priority used for review
Provide references using author date format
Wood, D. A. (2022), "Accurate brain‐age models for routine clinical MRI examinations." Neuroimage, 249, 118871.
Bashyam, V. M. (2020), MRI signatures of brain age and disease over the lifespan based on a deep brain network and 14 468 individuals worldwide. Brain, 143(7), 2312-2324.
Biondo, F. (2021), "Brain-age predicts subsequent dementia in memory clinic patients." medRxiv, 2021-04.