Weight Encode Reconstruction Network for Computed Tomography in a Semi-Case-Wise and Learning-Based Way

Classic algebraic reconstruction technology (ART) for computed tomography requires pre-determined weights of the voxels for the projected pixel values to build the equations. However, such weights cannot be accurately obtained due to the high physical complexity and computation resources required. In this study, we propose a semi-case-wise learning-based method named Weight Encode Reconstruction Network (WERNet) to co-learn the target voxel values and intrinsic physics of the case in a self-supervised manner without labeling the target voxel set. With the help of gradient normalization, the WERNet reconstructed voxel set with a high accuracy and showed a higher capability of denoising compared to the classic ART methods. Moreover, the encoder of the network is transferable from a voxel set with complex structures to unseen cases without the deduction of the accuracy. Our method can be applied in tomography-related applications and similar inversion problems even with unclear intrinsic physics.