Confounding Factors Mitigation in Brain Age Prediction Using MRI with Deformation Fields

The aging brain is characterized by a decline in physical and mental capacities and susceptibility to neurological disorders. Magnetic resonance imaging (MRI) has proven to be a critical tool in detecting age-related structural and morphological changes. One of the effective biomarkers for healthy aging is the difference between predicted brain age and chronological brain age (Predicted Age Difference: PAD). While deep learning networks give accurate brain age predictions, they are highly affected by confounding factors like variable noise levels from different MRI scanners/sites, gender, etc. This study focuses on the development of an algorithm leveraging deformation fields with debiasing in T1-W MRI images obtained from the OpenBHB dataset to learn representation vectors that capture the biological variability (age). To achieve this, we explore the use of learnable deformation fields combined with a contrast invariant training method (SynthMorph). We evaluate the accuracy of our method on the large publicly available dataset, OpenBHB, which consists of MRI scans from multiple sites and scanners and compare it with a standard available method, ResNet.