Image Enhancement in ASL Perfusion Imaging From Markov Random Fields to Variational Networks

Arterial spin labeling (ASL) perfusion imaging is a non-invasive technique capable of measuring the cerebral blood flow. Due to its poor signal-to-noise ratio, an efficient denoising method is required. Recently proposed methods like spatio-temporal total generalized variation (stTGV) improve the image quality but utilize full optimization procedures and thus are slow in inference. In contrast, deep learning (DL) based methods are fast in inference, but need much data and time for training. The aim of this thesis is to implement a co-sparse analysis model (CSM) and a variational network (VN) for ASL denoising to avoid both limitations.
A CSM uses learned filter kernels and applies a penalty function to the response of the filter. The result of this procedure forms a regularization term. In combination with a data fidelity term an "energy" is obtained which is topic to minimize w.r.t. an input image. In the framework of VNs, full optimization is replaced by an unrolled gradient descent scheme with a fixed number of steps. By using learnable penalty functions and a single parameter set for each descent step, a highly expressive and efficient model is obtained. Both models were trained with ASL data from 6 subjects and compared to stTGV on a quantitative and visual basis.
Although both models showed very good denoising performance, the VN outperformed the CSM. Despite visual differences, the VN and the stTGV performed on par in terms of structural similarity. However, the VN was about 50 times faster in denoising than stTGV. Further, the training of the VN lasts only 15 minutes.
This thesis highlighted the efficient ASL denoising capability of the VN. Its fast training and the ability to deal with few data makes the VN highly suited for more advanced applications in the field of arterial spin labeling.