| Diffusion Magnetic Resonance Imaging(d MRI)is a non-invasive imaging technique that detects the diffusion displacement distribution of water molecules in brain tissue to reflect changes in tissue microstructure.It is currently the only technology for detecting the integrity of white matter microstructure without invasion.High Angular Resolution Diffusion Imaging(HARDI)measures the diffusion of water molecules by acquiring diffusion-weighted images(DWIs)in hundreds of gradient directions to accurately reflect complex microstructures such as fiber crossings.However,due to the large number of gradient directions results in long scanning times,HARDI is limited in clinical applications.To alleviate this problem,some studies have used deep learning methods for fast reconstruction of HARDI data,but often only perform compressed sensing reconstruction in k-space or q-space,without fully utilizing the joint k-q sparse properties of HARDI data.In addition,supervised deep learning requires a large amount of data while it is difficult to obtain HARDI data.To address these issues,this thesis proposes the following semi-supervised deep learning model for fast high angular resolution diffusion imaging:(1)We design a k-q space joint compressive sensing and reconstruction algorithm based on order-aware uncertainty minimization(OAUM),which can reconstruct undersampled HARDI data in k-q space to fully sampled data.OAUM mainly consists of a k-space reconstruction module and a q-space reconstruction module.The k-space reconstruction module is mainly composed of spatial flip concatenate(SFC)and Kspace flip concatenate(KFC)modules,which enrich the features of spatial domain and frequency domain,and transfer information between the two domains using Fourier transforms,allowing the module to effectively recover the signal lost in k-space..The q-space reconstruction module uses multi-scale convolution blocks(MSCB)to provide a larger receptive field for the q-space reconstruction module,improve its performance by capturing multi-scale features.Finally,semi-supervised training of networks is achieved by cascading the k-space reconstruction module and the q-space reconstruction module in different orders and constraining the consistency of the reconstruction results of networks with different orders.Compared with cutting-edge fast-reconstruction methods,this method achieves the best results on DWIs,fiber orientation and corresponding diffusion parameter maps,proving that this semisupervised learning method can effectively perform fast HARDI reconstructions.(2)In order to reduce the impact of noise in HARDI data on the semi-supervised learning model for HARDI data reconstruction,this thesis proposes a DenoisedOAUM(DN-OAUM)model based on the semi-supervised learning model OAUM.We introduce a noise separation module based on graph framelet transforms(GFTs)into OAUM,which uses GFTs to filter the q-space signal and remove noise by representing the q-space signal as graphs,and then inputs the obtained denoised under-sampled HARDI data into OAUM network for reconstruction.Experimental results show that DN-OAUM outperforms OAUM in various reconstruction effect when reconstructing with noisy data. |
