Brain Extraction of Neurological FLAIR MRI for Multicenter, Multi-Disease Datasets

Fluid-Attenuated Inversion Recovery (FLAIR) MRI has emerged as an important sequence for analyzing neurodegenerative diseases. Large-scale, automated cross- sectional and longitudinal cerebral biomarker analysis from FLAIR datasets could progress disease characterization and help determine optimal intervention times. Brain extraction is a critical preprocessing step to such analytical pipelines. Despite this, most automated brain extraction algorithms are designed for T1-weighted or multi-modal inputs. This thesis presents the development of a deep learning-based brain extraction tool designed specifically for FLAIR. The proposed method, a Multiple Resolution U-Net (MultiResUNet), obtained mean Dice Similarity Coefficient (DSC) scores exceeding 98% on multicenter, multi-disease datasets. This enabled downstream clinical analysis where FLAIR biomarkers statistically differentiated (p<0.01) healthy, Mild CognitiveImpairment(MCI), andAlzheimer’s Disease (AD) patients. This work demonstrates that FLAIR alone can be used for end-to-end analysis of large datasets, which can lower acquisition costs, simplify clinical translation, and reduce measurement error associated with multi-modal approaches.