Motion artifact reduction in cardiac MR cine images based on deep-learning model (RF_SA_471)

This study aimed to develop a motion artifact reduction of cardiac magnetic resonance (CMR) cine images using noise removal based on the diffusion, and to examine the efficacy of the model by applying to the images corrupted by motion artifacts.

Methods:

For this study, expert radiologists chose 14 cardiac short axis (SAX) cine images with relatively low artifacts. The MR images were obtained by 3.0T scanner (Siemens). Each image was converted into the k-space by Fourier transform, and we artificially generated the artifacts based on the corresponding domain. The artifacts were generated by combining mis-triggering artifacts and breathing artifacts (Figure 1), and we manipulated the strength of artifacts on each slice image to generate images with various artifacts. Using the artificially artifacted images as a training dataset, we trained the diffusion model by considering artifact-free images as a label dataset. The dataset consisted of 5800 training data and 480 validation data. The test data consisted of images containing both artificial artifacts and real artifacts. 

Using the artificially artifacted images as a training dataset, we trained the diffusion model by considering artifact-free images as a label dataset. The dataset consisted of 5800 training data and 480 validation data. The test data consisted of images containing both artificial artifacts and real artifacts, which were used to evaluate the performance of trained model.

Results:

Figure 2 describes the results of test data. By using the data unused for training, we can see that our model can even correct the relatively strong artificial cine artifacts similar to the ground truth. By applying our model to the real artifacted images without ground truth, we can see that our model can correct the distortion of myocardium due to the breathing artifacts (red arrows) as well as mis-triggering artifacts, and remove the noise in the myocardium.

Conclusion:

In this study, we proposed a motion artifact reduction method for CMR cine images based on the diffusion model. Our diffusion model is trained based on the artificially generated motion artifacts, which enables to resolve the image registration issue due to the time difference in image acquisition. Finally, our model is capable of correcting artifacts without the distortion of images.