To achieve efficient and accurate quantitative susceptibility mapping (QSM) reconstruction based on susceptibility-weighted imaging (SWI) for brain magnetic susceptibility biomarker measurements in clinical diagnosis and longitudinal monitoring of neurodegenerative disease.

QSM enables non-invasive quantification of tissue magnetic susceptibility, providing direct measurements related to iron content, calcium deposition, and other paramagnetic or diamagnetic substances in the brain. Due to the high time and computational cost of QSM reconstruction, in daily neurological practice and in the majority of prior cohort studies, imaging protocols have primarily included susceptibility-weighted imaging (SWI) alone, thereby missing the ability to provide quantitative evaluations of magnetic susceptibility.

5083 healthy subjects with brain MRIs (SWIs and QSMs) were randomly selected from UKBiobank dataset. Data preprocessing (including normalization, orientation alignment, and registration) was performed to enhance sample consistency. A Generative Adversarial Network (GAN) with a U-Net generator and a multi-layer convolutional neural network has been designed as the generative model architecture. The FID score and the slice-wise correlation coefficient are used to evaluate the results.

Compared with QSM from UKbiobank data, the U-Net GAN-reconstructed QSM achieved an FID score of 1.6031 ± 0.17 and a slice-wise correlation coefficient (p < 0.01).