Convolutional Neural Network Optimization and Parallel Compressive Sensing Algorithms for Accelerated MRI Reconstruction
Abstract
Magnetic resonance imaging (MRI) is a noninvasive imaging modality that produces high-quality images. One of the biggest challenges in MRI is the lengthy scan procedure which could lead to motion artifact and patient discomfort. Due to the physical and physiological limits, undersampling the signals in the k-space signals has been used to shorten the scan time. However, the undersampling of k-space data results in undersampling artifacts that require advanced reconstruction algorithms to compensate for the missed signals. Many reconstruction algorithms have been proposed to address this problem. Linear interpolations in parallel imaging (PI) techniques usually suffer from high noise-like interpolation artifacts, and compressive sensing (CS) reconstructions are usually blurred in high-order undersampling factors. In this study, we first introduce a hybrid CS-PI algorithm and show it outperforms CS or PI individually in image reconstructions using actual data from MR-guided radiotherapy. Nevertheless, PI, CS, and hybrid CS-PI highly depend on the number of ACS in the center of the k-space and require a particular sampling strategy. Deep learning models can solve these problems with lower scan and reconstruction time with fewer interpolation artifacts and blurriness. In deep learning-based MRI reconstruction methods, the network’s architecture plays a crucial role in the quality of the reconstructed image. According to the large number of architectures that can be considered for these models, manually designing architectures and testing all the possible solutions are not practical. We introduce a new evolutionary-based search strategy to design a deep network for MR reconstruction automatically. We use different numerical metrics to compare the results of the optimized model with the ad-hoc model and three different published methods. The results showed that the proposed algorithm could automatically design a network that is not limited to only one particular sampling strategy and outperforms three related published models.
Citation
Vafay Eslahi, Samira (2022). Convolutional Neural Network Optimization and Parallel Compressive Sensing Algorithms for Accelerated MRI Reconstruction. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /198113.