Kantitatif manyetik duyarlılık haritalaması tekillik düzeltmesi için bir derin öğrenme yöntemi

Brain susceptibility maps have been used clinically to study the progression of neurodegenerative diseases such as Alzheimer's disease, Parkinson's disease and Multiple Sclerosis. One way of obtaining these brain susceptibility maps is through Quantitative Susceptibility Mapping (QSM). QSM is an inverse problem that aims to calculate susceptibility from MRI phase images. This is possible because the MRI phase image is, up to a constant, a convolution of a dipole kernel filter with an underlying susceptibility distribution. The problem is an inverse problem because the 3D dipole filter has zeros in the Fourier domain, which means direct inversion is impossible. Recently, deep learning-based solutions have been proposed to solve this problem. In this work, using synthetically generated data, we first examine the UNet deep learning architecture which has been the default architecture in the QSM community. We note that using a different variant can speed up training convergence with up to a 21% reduction in validation mean absolute error. Furthermore, we propose a QSM reconstruction method that solves the QSM problem by dividing the problem into two parts in the Fourier domain. On the one hand, we solve for susceptibility values by direct division in the Fourier domain in regions where the dipole kernel is stable. On the other hand, for regions where the kernel is zero, we train a deep learning model to learn the correct susceptibility map. Our final map is obtained by merging these two maps. We show that our reconstruction method works by testing on a generated test set. Our approach was also tested using data from a QSM challenge. On this data, we achieved the top position when compared with the other models tested during that challenge.