Application of A Deep Learning Strategy for Direct Estimation of Quantitative Susceptibility Maps

When compared with healthy individuals, there is an increase in the amount of iron deposit in the brain in patients suffering from diseases such as Parkinson's and Alzheimer's. Quantitative Susceptibility Mapping (QSM) is the key imaging method for quantifying iron deposit in the brain. While there have been traditional methods used for obtaining susceptibility maps from phase images, deep learning methods have recently been used for this purpose. One of the main deep learning architecture types used for obtaining QSM from phase images is the UNet. In this work, we modify this architecture and measure performance compared to standard approaches used in QSM literature. We found that, after training for the same number of epochs, our modified version obtained a 21% error reduction, suggesting that this modified version should be preferred to the base version.