Results

Evaluation

Three evaluation metrics were used to quantify the performance of our models. Two of them reflect the accuracy of CT numbers compared to fully sampled images: relative root mean squared error (rRMSE) and mean absolute error (MAE). The third metric measures the similarity between the images in terms of perceived quality: structure similarity index metric (SSIM). The SSIM metric takes into account the structural information present in the images, such as pixel correlations, brightness, and image contrast. To highlight the importance of integrating information from sinogram data, we compared the results achieved by the dual-domain CNN with those from the purely image-domain model.

Generalizability testing

Our current model was trained on the 140 kV dataset without exposure to any CT data acquired at other tube potentials. It is crucial to test the model generalizability since the conventional FBP and compressed sensing approaches could achieve consistent results across various acquisition scenarios. In addition to the tube potential, we also tested the model capability in handling even fewer view angles (61 views). Since the model was not trained on a 61-view dataset and fewer view angles made the sparse-view problem even more challenging, it is reasonable to expect that our model may not attain satisfactory performance in this scenario.

Results

The images reconstructed under different undersampling scenarios are illustrated in Figure 4. When the number of acquired view angles are doubled (i.e., 246 views), both image-domain and dual-domain results are free of streak artifacts that are present in the FBP images (shown in the first row of each case). This artifact removal process did not over-correct the images, causing shift of CT numbers or loss of spatial resolution. In scenarios where the acquired data is only 1/16th of the entire data set, severe artifacts render the anatomical structures nearly indiscernible in conventional FBP images. Both methods are able to diminish streaks and restore the structures, but the residual artifacts still persist. Despite the dual-domain method not achieving success in this more challenging scenario, it demonstrates greater capability in terms of reducing streak artifacts and preventing the reconstructed images from being excessively smooth.

Conclusion

In this work, we present a dual-domain method that initially corrects the artifact-contaminated images and then proceeds with a sinogram-domain model to enhance the accuracy of the reconstructed images. This approach leverages both spatial information and partially sampled sinogram data, demonstrating a greater capacity to reduce streak artifacts and restore anatomical structures than a pure image-domain model. Our method achieves consistently better results across varying undersampling scenarios. However, residual artifacts persist when the acquired data is only 1/16th of the complete data set. In future studies, we will explore a more efficient approach to handle such challenging tasks by incorporating the recently proposed and highly effective diffusion-based methods.