An increasing use of computed tomography (CT) in modern medicine has raised radiation dose issues. Strategies for low-dose CT imaging are necessary in order to prevent side effects. Among the strategies, limited-angle CT scans are being used for reducing radiation dose. However, the limited angle scans cause severe artifacts in the images reconstructed by using conventional reconstruction algorithms, such as filtered back-projection (FBP), due to insufficient data. To solve this issue, various methods have been proposed to replace conventional reconstruction algorithms. In this study, we proposed a data-driven deep learning-based limited-angle CT reconstruction method. The proposed method, called Recon-NET, consisted of 3 fully connected (FC) layers, 5 convolution layers and 1 deconvolution layer, and the Recon-NET learned the end-to-end mapping between projection and reconstructed image data. The FBP algorithm was implemented for comparison with the Recon-NET. Also, we evaluated the performance of the Recon-NET in terms of image profile, mean-squared error (MSE), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). The results showed that the proposed model could reduce artifacts and preserve image details comparing to the conventional reconstruction algorithm. Therefore, the Recon-NET has the potential to provide high-quality CT images with low-dose and reject the complexity of conventional techniques.
Stationary inverse-geometry digital tomosynthesis (s-IGDT) has advantages in terms of motion artifact reduction and diagnostic efficiency improvement. However, truncation artifacts are caused in reconstructed images owing to the geometric characteristics of s-IGDT systems, and this drawback degrades the diagnostic accuracy. In order to overcome this limitation, we proposed a convolutional neural network (CNN)-based truncation artifact reduction method. We simulated a s-IGDT system with stationary X-ray source array and small detector. Also, we acquired s-IGDT images using 70 volumetric phantoms based on the SPIE-AAPM lung CT challenge dataset. The U-Net was used as the CNN architecture, and we trained the network through 207 s-IGDT images. We confirmed that the truncation artifacts with various patterns included in the prior images were clearly removed in the prediction images obtained by the trained network. Moreover, the quantitative evaluation showed that both of the structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) improved when using the proposed method. The averaged SSIM and PSNR of the prediction images were approximately 6% and 25% higher than those of the prior images, respectively. In conclusion, the proposed model based on the CNN has superior performance in removing the truncation artifacts of s-IGDT images.
Digital tomosynthesis (DTS) has been used in diagnosis and radiation therapy due to its performance and benefits. DTS is able to provide 3D images with good depth resolution comparing to conventional radiography and reduce radiation dose comparing to computed tomography (CT). However, DTS scans with limited scan angles and a few projections lead to the insufficiency in data acquisition. Such a drawback causes the alteration of image characteristics and the distortion of image quality in DTS imaging. These issues can be magnified by the geometric variations of the DTS systems and the imaging strategies. In this study, the effect of geometric variations on image characteristics was evaluated by the DTS systems simulated with various X-ray source scan trajectories and angles. The DTS images were analyzed in terms of noise property, contrast-to-noise ratio (CNR), and spatial resolution. The results showed that the quality of DTS images was highly dependent on X-ray source scan trajectories and angles, and the characteristics of DTS images varied according to their acquisition geometries. Therefore, the geometries and strategies for DTS imaging should be appropriately determined for optimizing their systems and applications.
According to an increased use of computed tomography (CT) in medicine, the risk caused by radiation exposure has been considered as one of the major issues. In order to reduce the risk, low-dose CT imaging has attracted attention. However, the low-dose CT imaging causes low spatial resolution (LR) and high noise in reconstructed images. Recently, deep learning-based models have shown a feasibility for reducing noise and improving spatial resolution. However, these models have the drawbacks such as complex structures, large sample size and computational costs. In this study, a simple denoising and super-resolution convolutional neural network (SDSRCNN) was proposed to overcome the limitations of conventional methods. Two networks were trained for the denoising and super-resolution imaging separately, and the trained networks were linearly combined as a single network with a simple architecture. In comparison with conventional methods, denoise-autoencoder (DAE) and super-resolution convolutional neural network (SRCNN) were also implemented. We evaluated the performance of the SDSRCNN in terms of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). The results showed that the proposed model could efficiently reduce noise and preserve spatial resolution information comparing the conventional methods. Therefore, the proposed model has the potential for improving the quality of CT images and rejecting the complexity of the conventional methods.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.