Repeatability and reproducibility of brain age estimates in multiple sclerosis for three publicly available models

Abstract

Accelerated brain aging is a marker of disease-related neurodegeneration in multiple sclerosis (MS). Artificial intelligence models, trained on healthy individuals, can estimate age from brain MRI scans, but the effects of technical variations between MR scanners and conditions on these estimates are currently insufficiently investigated. This study aims to determine the within-scanner repeatability and between-scanner reproducibility of the brain-predicted age difference (brain-PAD) across three brain age models.

30 people with multiple sclerosis and 10 healthy controls (mean age 44.2 ± 11.7 years and 39.2 ± 12.9 years, respectively), underwent six scans in a single day; a scan and immediate on a 3 T GE, 1.5 T Siemens and a 3 T Siemens MRI-scanner. Brain-PAD was determined using brainageR, DeepBrainNet and the MIDI-model from 3D T1w brain MRI-scans. Intraclass correlation coefficient (ICC) was used to quantify absolute agreement within-scanner (ICC-AA) and between-scanner consistency (ICC-C). Variance component analyses were used to determine the standard error of measurement (SEM) and the smallest detectable change (SDC).

Brain-PAD was higher for pwMS compared to HC when predicted with brainageR and DeepBrainNet, not when predicted with the MIDI-model. Within-scanner repeatability was excellent (ICC-AA>0.93) for all models. Between-scanner reproducibility was good to excellent (ICC-C>0.85) for brainageR and the MIDI-model, while DeepBrainNet, showed excellent between-scanner reproducibility for Sola vs. VIDA (ICC-C:0.97), but moderate for GE vs. Sola and for GE vs. Vida (ICC-C:0.63 and 0.61). Between-scanner SDC was 6.56 years for brainageR, 5.57 years for the MIDI-model and 22.65 years for DeepBrainNet.

Our findings demonstrated high repeatability of brain age estimates from the same scanner, but variable reproducibility across different scanners, irrespective of the brain age prediction model.