Medical image reconstruction with multi-level deep learning denoiser and tight frame regularization

As a fundamental task, medical image reconstruction has attracted growing attention in clinical diagnosis. Aiming at promising performance, it is critical to deeply understand and effectively design advanced model for image reconstruction. Indeed, one possible solution is to integrate the deep learning methods with the variational approaches to absorb benefits from both parts. In this paper, to protect more details and a better balance between the computational burden and the numerical performance, we carefully choose the multi-level wavelet convolutional neural network (MWCNN) for this issue. As the tight frame regularizer has the capability of maintaining edge information in image, we combine the MWCNN with the tight frame regularizer to reconstruct images. The proposed model can be solved by the celebrated proximal alternating minimization (PAM) algorithm. Furthermore, our method is a noise-adaptive framework as it can also handle real-world images. To prove the robustness of our strategy, we address two important medical image reconstruction tasks: Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET). Extensive numerical experiments show clearly that our approach achieves better performance over several state-of-the-art methods.