Noise-Generating-Mechanism-Driven Unsupervised Learning for Low-Dose CT Sinogram Recovery

Deep learning (DL) techniques have expedited successful applications in computed tomography (CT) imaging field and have obtained remarkable outcomes. Most of the techniques require the precollection of large numbers of low-dose/high-dose CT image pairs and the training of the network in a supervised end-to-end manner. However, there are some evident limitations to such a supervised learning manner: 1) it is time intensive to acquire sufficient training sample pairs, as expected, and the low-dose CT images are usually simulated from the high-dose images due to the radiation exposure to the patients. In addition, it is difficult to obtain well-registered data pairs, especially for the torso region and 2) most current methods fail to utilize the abundant information in the unlabeled low-dose CT images, which are easier to obtain. In this work, we develop a new method for low-dose CT sinogram recovery with a noise-generating-mechanism-driven unsupervised DL network (Un-SinoNet), wherein the unlabeled low-dose CT sinograms are enrolled into the network training. The proposed Un-SinoNet considers the statistical variations in the noise-generation mechanism of CT measurement and its prior characteristics under a maximum
a posteriori
probability (MAP) framework, and then learns the proper gradient of the low-dose CT sinograms in a purely unsupervised manner. Specifically, similar to the labeled high-dose CT sinogram, the gradient information in the unlabeled CT sinogram can be used for efficient network training. Moreover, with a few low-dose/high-dose sinogram pairs, the proposed Un-SinoNet can be extended to the semi-supervised DL network (Semi-SinoNet) by combining the conventional objective function and the MAP objective function. The proposed Un-SinoNet and Semi-SinoNet are validated via the LUNGMAN phantom, and simulated Mayo Clinic patient datasets. With the two methods, this study achieves better image resolution and noise removal effects than the competing methods, and the presented Semi-SinoNet produces similar or even better performance compared with the supervised DL network.