Introduction

Contrast enhanced MR angiography (CE MRA) scans require longer acquisition time to fully reflect contrast agent bolus arrival and peak at the area of interest¹. Due to this, many artifacts including motion and noise may affect the acquired CE MRA data. Conventional motion correction methods such as pairwise intensity-based registration often exhibit errors due to signal intensity changes caused by contrast agent arrival and motion itself². Deep learning methods have shown good generative ability in the recent years³. However, precise correction of the CE MRA scans may not necessarily guarantee a high quality angiogram. In this study, we propose deep learning methods to correct motion and improve SNR on CE MRA data. Additionally we propose an angiogram-aware training approach to assure the accuracy of the resulting angiogram.

Method

CE MRA data from 60 patients, split to 50:10 for training and test, with varying degree of ischemia were used in this study. The CE MRA data for carotid arteries were acquired using a head/neck coil with a 3D TWIST sequence and the following parameters: TR/TE = 2.62/0.95 msec, matrix = 320 x 240 x 144, field of view = 300 x 400 x 172.8 mm3, scan direction = coronal, and number of dynamic scans = 28. The brain area was isolated by cropping the CE MRA data to the size of 140 x 150 x 144 and changing the view to axial. Only the cropped CE MRA data were used throughout this study.
Motion and noise corruptions were synthetically introduced to the CE MRA data. Two types of motion artifacts were introduced: translational and rotational. The translational motions were added by adding linear phase into central phase encoding lines (more often) or periphery (less often) to reflect the nature of TWIST-type keyhole acquisition. The rotational artifacts were introduced by rotating the CE MRA data at random time points at rotation angles between -5 to 5 degrees. The noise corruptions were added in random time points and the added Gaussian noise ranges from 0 to 10% of peak CE MRA signal.
The deep learning architectures used to perform correction were U-Net and U-Net + generative adversarial network (GAN). For both architectures, the input and output were modified to 28 channels reflecting the number of dynamic scans. Both architectures were trained twice, with and without the angiogram-aware loss function. The angiogram-aware loss function is specified as the difference between the maximum intensity projections (MIP) of the temporally subtracted CE MRA scans and the MIP of the temporally subtracted network output. The loss functions used in all training schemes were L1loss + optional angiogram-aware loss for U-Net and L1loss + optional angiogram-aware loss + binary cross entropy (BCE) loss for U-Net + GAN. Other training parameters were: batch size = 8, training epochs = 500, learning rate = 1e-4, and ADAM optimizer.
The correction accuracy was evaluated using peak signal to noise ratio (PSNR) and structural similarity (SSIM) of the CE MRA dynamic scans and the resulting final angiogram.

Results

Sample CE MRA dynamic scans before and after deep learning based artifact correction showed successful correction of the dynamic scans and close similarity to the ground truth (Figure 1). The deep learning results also showed matching bolus arrival and bolus peak time, which may hold clinical relevance. Furthermore, deep learning networks trained with angiogram-aware loss showed better similarity in terms of the angiogram than those without angiogram-aware loss (Figure 2). This was consistent with quantitative analysis, showing the networks using angiogram-aware losses yielded improvement in angiogram for both U-Net and U-Net + GAN methods (Table 1). The angiogram-aware loss improved dynamic scans PSNR by up to 7% and angiogram SSIM by up to 17% when used in training.

Discussion

The proposed deep learning methods have successfully corrected synthetic artifact corruptions in CE MRA data and produced accurate angiogram when trained with an additional angiogram-aware loss. The improvement in angiogram accuracy highlights the need to consider the final processing steps when performing artifact correction in a dynamic data. The corruption cases used in this study can be extended to reflect MRA-specific artifacts such as pseudo-stenosis, venous contamination, or ringing artifacts. Correction methods for all of these cases may reduce the need for repeated scans in CE MRA studies.

Conclusion

In this work, we proposed deep learning methods to perform motion and noise correction in CE MRA data. Additionally, we proposed angiogram-aware loss that improves the accuracy of the corrected angiogram even further. We believe this approach can serve as a solution for correction of corrupted CE MRA data.

Acknowledgements

No acknowledgement found.