Self-supervised deep convolutional neural network training for low-dose CT reconstruction

Abstract

Computed tomography (CT) is a medical imaging technique to obtain a noninvasively three-dimensional image of the body. CT imaging is one of the most crucial tools which is used for monitoring the human body and diagnosing serious illnesses. In CT imaging, one of the most serious concerns has been ionizing radiation since exposure to large amounts of radiation can cause serious illnesses. Various low-dose CT reconstruction approaches have been proposed to reduce the dose level without compromising image quality. With the emergence of deep learning, the increasing availability of computational power, and huge datasets, data-driven methods have recently gotten a lot of attention. Deep learning-based methods have also been applied in various ways to address low-dose CT reconstruction problem. However, the success of these methods is usually dependent on labeled data, which requires tedious work by radiologists for CT imaging. Recent studies, however, have also shown that training may be done successfully with noisy datasets without the requirement of noise-free target data. In this study, a training scheme is defined to use low-dose projections as their own training targets. We apply the self-supervision principle in the projection domain where the noise is element-wise independent, which is a requisite for self-supervised training methods. The parameters of a denoiser neural network are optimized through self-supervised training. Experiments are done with both analytical and human CT data. Slices from deep lesion dataset for human CT data and ellipses dataset for synthetic data are used. To simulate low-dose settings, 64 views parallel beam geometry is used. The noisy projections are created with additive white Gaussian noise with 30, 33, and 37 dB SNR values. The proposed method is compared with FBP, SART, SART+TV, SART+BM3D, and the supervised FBP+U-Net method. The methods are compared quantitatively with PSNR and SSIM metrics, and the reconstructions are qualitatively assessed regarding background smoothness, the sharpness of the details, and the recoverability of the lesions with some visual examples. In the comparisons, it is shown that the proposed method outperforms both traditional and compressed sensing-based iterative reconstruction methods in the reconstruction of analytic CT phantoms and real-world CT images in the low-dose CT reconstruction task, both qualitatively and quantitatively. Besides, it produces comparable results with the supervised approach.