1.

INTRODUCTION

Computed tomography (CT) has been widely used for disease diagnosis in the medical imaging field due to its detailed observations of anatomical structures. However, the X-ray radiation exposure can increase a potential risk of cancers to the patients. To reduce the X-ray radiation dose, decreasing the number of projection views during CT scan, called sparse-view CT, can be implemented, whereas the severe streak artifacts are generated in the reconstructed images.

To overcome this, several iterative reconstruction (IR) methods, which iterativley optimize the both CT data fidelity and the image regularization terms, have been developed for sparse-view CT. For the prior of image regularization term, total variation^{1, 2} is often utilized. However, IR calculates the forward and backprojection during optimization, requiring a huge amount of computational cost. Moreover, it is not easy to balance the parameters of fidelity and regularization terms for various imaging tasks. Therefore, the usage of IR methods is limited in medical imaging field, where the real time applications are required.

Recently, deep learning-based approaches have shown promising results in sparse-view CT reconstruction. Inspired by convolutional neural networks (CNNs), the image-domain method^3–5 trains the spatial information of streak artifacts distribution and reduces the streak artifacts effectively, whereas it also often reduces the signals overlapped by streak artifacts together. To overcome this limitation, the hybrid-domain approach,^6–8 which contains both image and sinogram domain, not only learns the image prior but also utilizes the information of measured projection data during network training. From this, the hybrid-domain method improves the performance of streak artifact reduction while preserving the edge sharpness better than when only the imagedomain is used.

Although the above CNN-based methods showed good performance, they are fully-supervised approaches that require the paired full and sparse-view CT images, where the anatomical structures are identical. In practice, it is not feasible to acquire the dataset of those pairs from the patients due to the principle of ALARA (as low as reasonably achievable). In this reason, to acquire the full and sparse-view CT image pairs, most of methods generate the sparse-view CT images by a computer simulation using the full view CT data. However, if only the sparse-view CT data are given, it is impossible to produce full view CT images that are anatomically identical. This is because image reconstruction in sparse-view CT is an undetermined inverse problem that has no general solution.⁹

To tackle this problem, we propose a weakly-supervised learning for streak artifact reduction when the only unpaired sparse-view CT data are given. We acquire the pairs of data for CNN training from the given sparse-view CT data and train the network. We then apply trained network to the given sparse-view CT images iteratively. For the success of our framework, we perform streak estimation model for signal preservation.

2.

METHODS

Figure 1 shows the schematic diagram of the proposed method for reducing streak artifacts with unpaired sparseview CT data.

Figure 1.

The overall schematic of the proposed method. FP: forward projection; FBP: filtered back projection

At times it may be desired, for formatting reasons, to break a line without starting a new paragraph. This situation may occur, for example, when formatting the article title, author information, or section headings. Line breaks are inserted in LaTeX by entering\\or\linebreak in the LaTeX source file at the desired location.

2.1

Weakly-supervised learning

For CNN training to reduce streak artifacts, it is required to set the images with strong and weak streak artifacts as input and target, respectively. Since the only unpaired sparse-view CT data are given in our scenario, we should acquire pairs of the input and target from given sparse-view CT data. For this, we generate an image that has stronger streak artifacts than the original streak artifacts in given sparse-view CT image, which is denoted as a sparsier-view CT image. We then regard the sparsier and sparse-view CT images as the input and target for CNN training, respectively.

Here, the method of sparsier-view CT image generation is described in the following procedures. To make streak artifacts stronger, it can be achieved by decreasing the number of projection views because the strength of the streak artifacts is dependent on the number of projection views. Therefore, we down-sampled the given sparse-view sinogram by half and acquire a sparsier-view sinogram. Then, we reconstruct the sparsier-view CT image from the sparsier-view sinogram with the filtered backprojection (FBP) algorithm.¹⁰ Note that we set down-sampling ratio as 2 to ease CNN training because the spasier view CT image will be significantly corrupted when the down-sampling ratio is bigger than 2.

Let and x be the sparsier-view CT image and sparse-view CT image, respectively. Since the projection data of is extracted and down-sampled from that of x, the streak artifacts in will be amplified while keeping the directionality of the original streak artifacts in x. Therefore, we set a pair of and x as CNN training dataset. We then train CNN with the mean squared error (MSE). The MSE is defined as

where f (·; θ) and θ denote the network operator and network parameters, respectively.

3.

DATASETS AND EXPERIMENTS

4.

RESULTS

Figure 2 shows the resulting images on XCAT dataset. The linear interpolation image showed less streak artifacts than 128 view FBP image through the estimation of missing view data, whereas the secondary artifacts such as edge distortion were produced due to the interpolation errors. The fully-supervised method reduced the streak artifacts and noise significantly. However, it can be seen that the lesions indicated by red and yellow arrows in ROIs of the fully-supervised method were blurred, leading to poor performance of visibility of those signals. In contrast, the proposed method produced the best visual similarity of ROIs to those of 512 view FBP image while preserving the edge sharpness. For the proposed method with iteration 1, its difference image had more edge errors than the results of other iterations. However, these errors disappear as more iterations are conducted.

Figure 2.

The reconstructed images with XCAT dataset. The display window of field of view (FOV) and zoomed region of interest (ROI) images is (-240,320) in HU. The display window of difference images is (-100,100) in HU.

Table 2 summarizes the average and standard deviation of normalized root MSE (NRMSE) and structural similarity index (SSIM)¹⁴ on XCAT testset. We observed that the fully-supervised method showed the best scores of NRMSE and SSIM. For the proposed method, iteration 2 had the lowest NRMSE and highest SSIM scores. Although the proposed method with iteration 2 showed slightly worse scores than the fully-supervised method, visual inspection confirms that the proposed method produces better image texture without sacrificing the detectability of lesions.

Table 2.

The quantitative evaluations of XCAT dataset

To examine the effect of the streak estimation model in the proposed method, we did not apply the streak estimation model after the iterative streak artifact reduction procedure and the results are shown in figure 3. As the number of iterations increases, although the streak artifacts are gradually reduced, it can be seen that the images were getting blurred. For the ROIs of more than iteration 2, the signals were lost and the edge was also over smoothed. Although the results of iteration 1 preserved the signal indicated by red arrow, the streak artifacts were not reduced enough. From this, we chose to use iteration 2 to effectively reduce streak artifacts while preserving edge and signal shapes.

Figure 3.

Prior images after conducting different number of iterations in the proposed method. The display window of FOV and ROI images is (-240,320) in HU. The display window of difference images is (-100,100) in HU.

5.

CONCLUSION

In this work, we proposed the weakly-supervised learning for streak artifact reduction with unpaired sparse-view CT data. From the results, the proposed method achieved the best performance in preserving lesions while reducing streak artifacts compared to the other results. By overcoming the difficulty of acquiring the pairs of sparse and full view CT images in practice, we expect that our novel framework can be utilized successfully in medical imaging field.