Performance of Deep Learning-based Image Denoising in Image Reconstruction for Various Acquisition Conditions: a Simulated Phantom Study
Parisa Asadi, Andriy Andreyev, Matthew Andrew · Microscopy and Microanalysis · 2023
The use of X-ray computed tomography (CT) imaging has revolutionized the ability to non-destructively inspect the internal structures and features of various objects. Despite its advantages, the reconstruction and generation of high-quality 3D CT images remains a challenging task that is computationally intensive, time-consuming, and dependent on the image analysis method. This study aims to evaluate the performance of a deep learning-based noise-to-noise image reconstruction method (DeepRecon Pro) over various acquisition conditions using Shepp-Logan phantom by digitally simulating the count statistics for various noise levels and numbers of projections. 3D X-ray Computed Tomography (CT) imaging is a powerful tool for non-invasive examination of objects and is widely used in medical, industrial, and scientific research fields. However, the acquired images often contain noise reducing the quality and accuracy of the 3D reconstructed images. One promising approach for CT image noise reduction is the noise-to-noise image denoising method using deep neural networks. The basic idea behind this method is that the noise in each image is uncorrelated, and by averaging multiple noisy images, the noise can be reduced while preserving the underlying signal. The DeepRecon Pro [1] is one of a kind of deep learning methods, that involves the training of a U-Net-based deep neural network, which takes projection images or reconstructed volumes as input and outputs improved projection images or improved reconstructed volumes, respectively. In this paper, we will evaluate the performance of deep learning-based noise-to-noise image reconstruction method (Deeprecon Pro) for accurate image reconstruction and compared with the results obtained from traditional methods (Figure 1). A three-dimensional artificial phantom (Shepp-Logan) was employed in this study. As Figure 1 shows, first, a Shepp-Logan phantom was forward projected using cone-beam circular geometry and a sinogram was obtained. Then, different levels of Poisson noise, modeling various X-ray exposure times, were added to the sinogram, repeated for a different number of projections, and modified sinograms were reconstructed. Next, the deep learning U-Net based model was trained by generating reconstructed volumes for two or more data subsets. After that, the trained model was applied to improve the image quality and reduce the noise level of the images in matching conditions. The reconstructed results were cross compared with the ground truth data (original phantom 3D image used for forward projection), the baseline FDK (Feldkamp, Davis and Kress) reconstructed images, and non-local means (NLM) filtered images. Figure 2 shows the performance of the models tested on datasets with various Poisson noise (higher Poisson noise value, less noise) for 2 times projection reduction. The models tested on the first pair of data (odd projections), second pair (even projections) and total projected images. The result showed (Figure 3) that the spurious signal can be effectively removed by the trained network which results in lower mean square error (MSE) and higher structural similarity index measure (SSIM) compared to the FDK and NLM denoised results. For example, for 1600 projections and same level of noise, MSE reduced from 0.457 to 0.003 and the SSIM increased 0.008 to 0.750, when using deep learning based denoising. The performance of models for reconstruction of the data with different number of projections (Figure 4) also showed the superiority of deep learning-based model over traditional methods of reconstruction. For example, increasing the number of projections from 1600 to 4500, improved the performance of traditional methods about two times (MSE reduced from 0.45729 to 0.16359) that demonstrates the dependency of these methods on more reliable data for correct reconstruction. On the other hand, these values improved only from 0.0033 to 0.0031 for deep learning method, showing that much lower amount of data is needed. These results were consistent for different level of noise and different number of projections which shows the robustness of the deep learning-based noise-to-noise image reconstruction model over FDK or traditional filtering models. This study provides valuable insights into the use of deep learning-based image reconstruction methods in CT imaging. The results show that deep learning-based image reconstruction methods are an effective and efficient solution for improving the quality of CT images in various acquisition conditions. In addition, it provides a promising solution to address many issues related to the image quality versus acquisition time trade-off. Model training workflow, the reconstructed images were compared with 3D phantom was considered as ground truth data. Performance of different image processing methods for different level of noise, higher the Poisson noise value, less noise added. Visual performance of the deep recon (DR) model tested on odd pair (DR-1), second pair (DR-2), and total projections, versus traditional methods (for 1500 Poisson noise level and 1600 total number of projections). Performance of image processing methods as function of number of projections.