Background: MRI is a powerful medical imaging modality that provides excellent contrast between soft tissues. However, MRI signal acquisition and image reconstruction usually takes more time compared to computed tomography and ultrasound. In some applications such as real time imaging, it is important to acquire the MR signal and reconstruct the image with minimal latency. Practical approach to accelerate MRI is acquiring under-sampled k-space and reconstructing the image. Compress sensing methods is the state of the art approach to reconstruct the image from heavily under-sampled k-space. However, in CS-MRI, finding a proper regularization is time consuming and the image reconstruction algorithm is typically
associated with long reconstruction time beyond acceptable for real time imaging applications. Recent advances in deep neural networks opens new possibility to ease the reconstruction problem in an efficient manner. In this work, we propose to use the deep CNN to reconstruct the real-time 2D short axis cine images from regular under-sampled raw data.
Method: free breathing fast cardiac fully sampled images were acquired from five healthy volunteers. Raw k-space data of each receiving channel is sub-sampled and combined together by applying square root of sum of squares. The first three volunteers data were used in training the neural network, the fourth volunteer was used as a validation set and the fifth volunteer was used to test the network. Also, three different loss functions for the training step were tested: L2, L1, Structural Dissimilarity Loss Function( DSSIM). Different under-sampling rates were evaluated to determine maximum achievable rate. What’s more, we modify a real time MRI pulse sequence to acquire sub-sampled k-space to show the feasibility of prospective dynamic imaging.
Results: Based on our preliminary results, for regular under sampling pattern, 8 folds is the maximum achievable rate. A part of the left-ventricle’s border vanished when training the network for increased regular under-sampling. Higher acceleration rates could possibly be achieved by implementing prior knowledge or changing to an irregular under-sampling pattern. Among the mentioned loss functions, DSSIM loss function has the lowest Peak Signal to Noise Ratio (PSNR) and highest SSIM.
For prospective study, results for 4X and 6X acceleration were promising. The reconstruction time is around 6 ms for each image.
Conclusions: Deep learning can potentially shorten the MR acquisition time and decrease the reconstruction time, making it suitable for real time imaging. For clinical application, training the network on large clinical dataset would be necessary.