Quantitative Comparison of Deep Learning-Based Image Reconstruction Methods for Low-Dose and Sparse-Angle CT Applications

Abstract

The reconstruction of computed tomography (CT) images is an active area of research. Following the rise of deep learning methods, many data-driven models have been proposed in recent years. In this work, we present the results of a data challenge that we organized, bringing together algorithm experts from different institutes to jointly work on quantitative evaluation of several data-driven methods on two large, public datasets during a ten day sprint. We focus on two applications of CT, namely, low-dose CT and sparse-angle CT. This enables us to fairly compare different methods using standardized settings. As a general result, we observe that the deep learning-based methods are able to improve the reconstruction quality metrics in both CT applications while the top performing methods show only minor differences in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). We further discuss a number of other important criteria that should be taken into account when selecting a method, such as the availability of training data, the knowledge of the physical measurement model and the reconstruction speed.

Improving reproducibility in synchrotron tomography using implementation-adapted filters

“…When insufficient training pairs are available, various unsupervised approaches, such as the Deep Image Prior method proposed by Ulyanov et al (2018), are more suitable. For a quantitative comparison of various popular deep-learning-based reconstruction methods, we refer the reader to Leuschner et al (2021).…”

Efficient Strike Artifact Reduction Based on 3D-Morphological Structure Operators from Filtered Back-Projection PET Images

“…Finally, STD, FWHM, and SNR were used to evaluate denoising and image quality after noise processing. In addition, when 3D-MSO was used to process the FBP nuclear medicine images with the Deluxe Jaszczak phantom, a large number of operator combinations (for instance, morphological structural operator combinations of 3 × 3 × 3 had 2 27 , i.e., 134,217,728, combinations) required an extremely long duration when the operation was performed for each image. Therefore, the optimal response curve (ORC) was used to reduce the number of large datasets to 200 or less, and then 3D-MSO was used to identify the operator combinations with the lowest background noise for the images.…”

“…The number of operator combinations were 2 8 (256), 2 12 (4096), 2 12 (4096), 2 18 (262,144), 2 12 (4096), 2 18 (262,144), 2 18 (262,144), and 2 27 (134,217,728). The principle of matrix convolution was used to perform the operation in the morphological structure matrix.…”

“…Moreover, the convolutional neural network (CNN) methods were incorporated with quantitative study under the conditions between low dose and sparse matrix on CT images. The experimental results were shown and were demonstrated to be able to improve the quality of noise images [27]. Regardless of the successfulness obtained by CNNs on CT images, the accuracy of quantitation in PET images is still to be confirmed [28,29].…”

Deep Learning in Medical Image Analysis

“…Leuschner and Schmidt (2021) [ 1 ] from Germany, the Netherlands, and Canada present the results of a data challenge that the authors organized, bringing together algorithm experts from different institutes to jointly work on quantitative evaluation of several data-driven methods on two large, public datasets during a ten-day sprint.…”