Estimating brain age in MS: machine learning versus deep learning

A Norwegian study compared traditional machine learning with deep learning methods to estimate brain age in a large, longitudinal cohort of people with multiple sclerosis (MS). Both brain age models revealed significant associations with Expanded Disability Status Scale (EDSS) score and duration of disease. The authors concluded that deep learning is an ideal analytic tool to process complex imaging data.

Brain age is a numerical estimate of biological age, which is calculated by combining structural T1 MRI images and information about chronological age, in conjunction with artificial intelligence-based computation [1]. Dr Lars Skattebøl (Oslo University Hospital, Norway) aimed to validate a deep learning simple fully convoluted neural network (DL-SFCN) model for brain age in MS. The main objective was to compare DL-SFCN with an established feature-based machine learning model. To this end, eligible MRI data of 1,516 MS patients was analysed [2]. Since there is no “true” comparative value for a brain age estimate, Pearson's correlation and linear mixed effect (LME) modelling were used to look for associations between brain age, age, and clinical variables.

In deep learning estimations, brain age and chronological age were more firmly correlated (correlation coefficient [R]=0.90; 95% CI 0.89–0.90) than in machine learning estimations (R=0.75; 95% CI 0.74–0.76). Higher brain age was significantly associated with higher EDSS for deep learning estimates (t=5.3; 95% CI 0.17–0.37) as well as machine learning estimates (t=3.7; 95% CI 0.16–0.51), and with longer disease duration for deep learning estimates (t=5.8; 95% CI 0.08–0.15) and for machine learning estimates (t=6.5; 95% CI 0.15–0.28). The strongest correlation was found using longitudinal data. Clinical correlation between brain age and brain age gaps (BAG) was stronger at lower EDSS scores and was non-significant at an EDSS score of 6. The largest differences in BAG were found at an EDSS score of about 6, these differences were less prominent at lower EDSS scores.

Overall, the DL-SFCN performed equally well to the model comparator, both showing significant correlations to EDSS and disease duration. Deep learning may be of higher clinical value due to a stronger association to EDSS than machine learning.

1. Franke K, et al. Neuroimage. 2010;50(3):883–92.
2. Skattebøl L, et al. Brain Age in Multiple Sclerosis: A comparison of traditional machine learning and deep learning methods. OPR-123, EAN 2022, 25–28 April, Vienna, Austria.

Copyright ©2022 Medicom Medical Publishers



Posted on 22 August, 2022