Comparing 2D and 3D deep learning (DL)-based reconstruction strategies for under sampled isotropic 4D whole thoracic automated cardiac MRI (AutoCMR) (RF_SA_472)

One of the challenges facing 3D cine MRI is the difficulty of reconstructing such large volumes of data due to intractable memory and reconstruction time requirements. A single slice cine with matrix size 128 x 256 and 30 cardiac frames generally requires 220 MB of memory to reconstruct regardless of algorithm (compressed sensing (CS), deep learning (DL) based¹ etc.), easily fitting on modern GPUs. 3D cines can quickly exceed these hardware limits by 1000x and therefore require limiting reconstruction in the time-dimension or in space². DL inferencing has been shown to overcome these memory limitations and reduce reconstruction time. However, applying DL to 3D cines can be performed using different strategies, such as: (a) 2D spatial, (b) 2D+time or (c) 3D spatial. We evaluated these reconstruction strategies on undersampled isotropic 4D whole thoracic automated cardiac MRI (autoCMR) data and assessed ensuing image artifacts to find the optimal strategy for our data.

Methods:

We recently acquired multidimensional cardiac data including isotropic 3D cine data on a 3T MR System (Cima.x, Siemens Healthineers) using AutoCMR, a GRE-based free running acquisition. The acquired data was undersampled in K_x-K_y in a variable density sampling pattern, respiratory motion corrected, and resolved into 30 cardiac frames. 

We studied three models to find the optimal reconstruction for our acquisition by minimizing a weighted sum of Mean Squared Error³ and focal frequency loss⁴ for a) 2D UNET⁵ trained on 2D axial slices (Figure 1A), b) 3D UNET trained on 30 cardiac phases of 2D axial slices stacked as 3D volumes (Figure 1B) and c) 3D UNET trained on 3D spatial volumes (Figure 1C), trained using CS reconstructions. A data split of 72/9/9 for training/validation/testing was used on a single Nvidia A100 GPU considering maximum memory requirements (Figure 1D).

Results:

Figure 2A shows three orthogonal views of a dataset reconstructed using Fast Fourier Transform (FFT), CS and the models discussed above. The corresponding error maps are shown in Figure 2B. We noted that the spatio-temporal (2D+time) model and 3D spatial model outperformed the 2D model despite being trained on fewer instances (13824 instances, 2160 instances and 414720 instances, respectively).

The 3D spatial results have higher PSNR than the spatio-temporal results. To evaluate motion, the displacement maps in Figure 3 show points on a line passing through a short axis slice over time for all results. The maps from the spatial models were observed to be noisier than the 2D+time model which shows reduced motion.

Conclusion:

We demonstrated 2D+time and 3D spatial deep learning reconstruction methods have higher potential as a denoising solution for 4D undersampled cardiac MRI data despite being trained on fewer instances than the 2D spatial model. The 3D spatial and spatio-temporal models pose different temporal artifacts and can be chosen based on diagnostic application requirements.