Brain Age Estimation on T2-FLAIR Scans for Application to Multiple Sclerosis.

The T2-FLAIR-based Inception-ResNet-V2 model achieved accurate brain age predictions comparable to T1-weighted-based models.

• T2-FLAIR model test set MAE = 3.31 years, R2 = 0.944; 5x ensemble MAE = 2.81 years, R2 = 0.955
• T1-weighted model test set MAE = 3.34 years, R2 = 0.942; 5x ensemble MAE = 2.84 years, R2 = 0.955
• Comparison between T2-FLAIR and T1-weighted models showed no significant difference (p = 0.91)
• Models were compared using t-tests based on bootstrapped standard errors

T2-FLAIR-based brain-predicted age difference (brain-PAD) was significantly higher in people with multiple sclerosis (pwMS) than in healthy controls.

• Brain-PAD was 7.07 years in pwMS versus -0.50 years in healthy controls
• Difference was statistically significant (p < 0.0001)
• A linear model framework was used, adjusting for age and sex

T2-FLAIR-based brain-PAD correlated significantly with MS disease duration.

• Correlation coefficient R = 0.24, p < 0.0001
• Analysis was conducted adjusting for age and sex using a linear model framework
• This correlation was consistent with the T1-weighted brain-PAD findings

T2-FLAIR-based brain-PAD correlated significantly with physical disability as measured by the Expanded Disability Status Scale (EDSS).

• Correlation coefficient R = 0.30, p < 0.0001
• Analysis adjusted for age and sex using a linear model framework
• This finding was consistent with T1-weighted brain-PAD results for EDSS

Brain age predictions from T2-FLAIR scans were most strongly driven by subcortical regions, particularly the thalamus.

• Saliency maps were obtained using the SmoothGrad method to visualize regions most important for predictions
• Subcortical structures, with the thalamus being particularly prominent, were identified as key contributors
• This finding applied to the T2-FLAIR-based model predictions

The study used a multicentre cohort of healthy participants for brain age modeling and a single-centre cohort of pwMS and healthy controls for external validation.

• Both 3D T2-FLAIR and 3D T1-weighted brain MRI scans were collected
• The Inception-ResNet-V2 architecture was used for 3D convolutional neural network modeling
• Clinical validation assessed relationships with MS diagnosis, clinical phenotype, disease duration, and EDSS

Most existing brain age models rely on 3D T1-weighted scans, which are not routinely acquired in MS clinical practice, limiting their clinical translation.

• T2-FLAIR is described as 'the core sequence for MS diagnosis and monitoring'
• The reliance on T1-weighted scans for brain age modeling was identified as a barrier to clinical use in MS
• This motivated development of a T2-FLAIR-based brain age model