Image reconstruction using deep learning for 3D compressive sensed undersampling MRI

• Main Authors: Po-Hsiung Chang, 張博雄
• Other Authors: Cheng-Wen Ko
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/ujg44d

碩士 === 國立中山大學 === 資訊工程學系研究所 === 107 === Magnetic resonance imaging can collect anatomical and metabolic information in vivo, which can be used as helpful tool for clinical diagnosis. Accelerating MR acquisition time has been considering one of the key issues in order to gather more information within limited time. In this regard, compressed sensing is considered as a promising option. CS-MRI reduces scan time by acquiring k-space data less than as Nyquist sampling criterion, followed by sophisticated reconstruction algorithms to achieve satisfactory image quality. Recently, the interest in using deep learning to reconstruct images has been largely increasing. In this study, we will reconstruct CS-MR images with deep learning algorithms. Random and linear undersampling schemes of compressed sensing with acceleration rate of R=3 and R=5 were applied to full-sampled T1 3D MR images at both of 1.5T and 3T. Two deep learning frameworks (convolutional neural network and U-Net) were constructed for image reconstruction. In total of 48 models were trained and tested. Our results demonstrated that U-net can achieve better reconstruction in both of 2D and 3D architecture. Random sampling scheme of CS and the coincidence between training and testing images in field strength benefit the performance of reconstruction. However, the hardware limitation leads to longer training time for U-net, which may need further improvement for practical usage.