Research On Mojette Projection Acquisition In Computed Tomography Reconstruction

X-ray computed tomography(CT)has been widely utilized in clinical diagnosis,industrial inspection,social security and other applications.With the increasing use of CT scanners,complete data cannot be acquired in practical applications frequently due to the limitation of physical factors such as radiation dose,scanning geometry,system hardware,scanning time and others.However,in the case of incomplete data,reconstruction algorithms based on Radon transform in continuous domain face an ill-posed problem,thus often yield bad results with severe artifacts.As a discrete form of Radon transform,Mojette transform is exactly invertible even with few samples thanks to its unique properties,which points out a feasible way for reconstruction from incomplete projection data.However,Mojette transform acquisition is incompatible with physical CT scanners,and the conversion from Radon projection is the only way to obtain Mojette projection.That hinders applying Mojette transform into industrial application.To address the situation,the major research work is as follows:(1)This thesis applies deep learning techniques into Mojette projection acquisition for the first time.It is known that both Radon transform and Mojette transform could be considered as one-dimensional feature representations of a 2D image.Therefore,in the view of feature expression,Mojette projection generation from data acquired Radon projection data in continuous domain is equivalent to image features transformation.Inspired by this intuition,this thesis presents the technical route of employing deep learning to learn the mapping between Radon projection and Mojette projection,which provides a promising way to acquire Mojette projection in practice.(2)A method of Mojette projection acquisition based on feedback mechanism is proposed.In the noise-free environment,the mapping from Radon projection into Mojette domain can be achieved using the proposed method in equal space sampling condition.In this method,the presented network takes Radon projection as the input and outputs the corresponding Mojette projection.Besides,the network adopts error feedback to improve the accuracy.In the training process,specific initialization based on physical model is introduced as a kind of prior knowledge to speed up convergence.(3)A method of Mojette projection acquisition based on feature fusion is proposed.In the noisy environment,the mapping from Radon projection into Mojette domain can be achieved using the proposed method when Mojette sampling interval is smaller than Radon's.The network is also structured on the combination of error feedback and specific initialization.Furthermore,considering the peculiarity of this issue,the presented network designs multi-channel fusion module to suppress the noise,propose feature fusion block to enhance the learning ability and employs channel-wise attention module to achieve accurate error feedback.To train the proposed network adequately,this thesis collects plenty of clinical reconstructions and synthesizes simulation slices,thus an image dataset is produced.Based on the produced imaged dataset,a projection transformation dataset is established,which makes a base for future research.Exhaustive experimental results have confirmed the effectiveness of the proposed method.The mapping between Radon transform and Mojette transform can be established by the proposed method.In addition,the designed network architecture is universally applicable and the design idea is reliable.