Deep learning based image restoration algorithm for CT images from imperfect acquisition conditions

Abstract

This paper proposes deep learning based image restoration algorithms for CT images from imperfect acquisition conditions. X-ray computed tomography is one of the most widely used medical imaging modalities, which can acquire detailed cross-sectional images of a patient in a non-invasive manner in a short time. However, there are many cases in which CT measurements are obtained imperfectly for various reasons. The CT image reconstructed with imperfect measurement data has low image quality, and makes diagnostic errors. In this paper, we deal with two cases of improving CT images reconstructed with imperfect measurement data. Limited-angle X-ray computed tomography is often used in many medical applications. However, due to the incomplete projection data, the limited-angle problem is highly illposed, so the reconstruction images suffer from severe artifacts having a globally distributed directivity. Existing iterative methods for the limited-angle problems require extensive computation, so here we propose a novel deep learning architecture that provide accurate reconstruction instantaneously. More specifically, on the basis of the recent theory of deep convolution frameworks that show that the power of deep learning derives from the synergetic interplay between the non-local and local bases, the proposed deep learning architecture first decorrelates the global redundancies using redundant contourlet transform, after which the remaining non-local and local redundancies are exploited using the cascaded applications of the dual frame Unets. Experimental results confirm that the proposed method effectively eliminates the artifacts, thereby preserving edge and global structures of the image. Recently, deep learning approaches have been extensively studied for low-dose CT denoising thanks to its superior performance despite the fast computational time. In particular, cycleGAN has been demonstrated as a powerful unsupervised learning scheme to improve the low-dose CT image quality without requiring matched high-dose reference data. Unfortunately, one of the main limitations of the cycleGAN approach is that it requires two deep neural network generators at the training phase, although only one of them is used at the inference phase. The secondary auxiliary generator is needed to enforce the cycle-consistency, but the additional memory requirement and increases of the learnable parameters are the main huddles for cycleGAN training. To address this issue, here we propose a novel cycleGAN architecture using a single switchable generator. In particular, a single generator is implemented using adaptive instance normalization (AdaIN) layers so that the baseline generator converting a low-dose CT image to a routine-dose CT image can be switched to a generator converting high-dose to low-dose by simply changing the AdaIN code. Thanks to the shared baseline network, the additional memory requirement and weight increases are minimized, and the training can be done more stably even with small training data. Experimental results show that the proposed method outperforms the previous cycleGAN approaches while using only about half the parameters.