Utvärdering av en djupinlärningsmodell för hjärnåldersprediktion : Påverkan av MRI-fältstyrka, skannertillverkare och kliniskt stadium av Alzheimers sjukdom

Brain age prediction models, trained on magnetic resonance imaging (MRI) data, have become powerful research tools for assessing brain health and studying neurological diseases such as dementia. This study evaluates a deep learning-based brain age prediction model with respect to one of the primary challenges of such models, achieving robustness under varying real-world imaging conditions. More precisely, the evaluation focuses on how prediction errors of the model are influenced by varying scanner manufacturers and MRI field strengths, across healthy controls, individuals with mild cognitive impairment (MCI), and individuals with Alzheimer’s disease (AD). Ensuring reliable predictions across diverse imaging conditions is critical if the model is to be applied clinically, where variations are common. The brain age model was applied to data from the Alzheimer’s disease neuroimaging initiative, which includes diverse diagnoses, and follow-up image acquisition with different equipment. Analysis followed a two-fold approach. First, a subset analysis examined differences within subjects between scans taken with field strengths of 1.5 T and 3 T. Second, robust Bayesian linear mixed models were fitted for each diagnostic group to model the effects of scanner manufacturer and field strength on brain age predictions while accounting for longitudinal and individual-level variability. The results showed that both field strength and scanner manufacturer systematically influenced model predictions, and these effects varied across diagnostic groups. Additionally, the model predicted progressively higher brain ages for MCI and AD, consistent with disease-related brain aging. These findings highlight how varying MRI acquisition factors influence the model’s predictions and point to the importance of understanding these acquisition-related biases to improve the model’s robustness and generalizability.