1333960 - Deep learning-based Rapid Image Reconstruction for High-resolution Cartesian First-pass Myocardial Perfusion Imaging at 3 Tesla

First-pass contrast-enhanced myocardial perfusion imaging is useful for evaluating coronary artery disease¹. 2D single-slice (SS) Cartesian perfusion imaging using compressed sensing (CS)-based image reconstructions such as L1-SENSE² enables fast and high-resolution imaging, but whole-heart coverage cannot be achieved without simultaneous multi-slice (SMS) imaging and the reconstruction is typically performed off-line (~45 minutes per slice). To address these limitations, we developed a deep learning-based (DL) reconstruction technique for high-resolution Cartesian perfusion imaging at 3 Tesla (T), for both SS and SMS MB=2 acquisitions, which provides fast and high-quality reconstruction and makes rapid online reconstruction feasible.

Methods:

Figure 1-A shows the proposed physics-driven unrolled image reconstruction network, which consists of several denoising modules, and each module has a CNN-based denoiser and data fidelity update preserving fine features. 4 repetitive denoising modules were implemented which is the maximum capacity allowed in our GPU (NVIDIA Tesla A100, 40 GB memory) due to memory limitation. Shared weights were utilized for each denoising module. The inputs to the network are concatenated single-channel complex-valued image series after coil combination³. Detailed acquisition parameters are listed in Figure 1-B. SS and SMS have a total acceleration of 4 and 8 (4*SMS MB=2), respectively.

To test the performance, 25 perfusion data from 10 healthy volunteers and 5 patients undergoing clinically ordered CMR studies with gadolinium on 3 T scanners (SIEMENS Prisma/Skyra) were used. 30 slices from 10 subjects with SS acquisitions were used for training. Another 15 slices from 5 subjects with SS and 30 slices from 5 subjects with SMS acquisitions were used for testing. Prospective images reconstructed using both (SMS)-L1-SENSE and DL were graded by an experienced cardiologist (5, excellent; 1, poor).

Results:

For SS acquisitions, image quality scores were 4.5 [4 ,5] and 3.6 [3.5, 4] for L1-SENSE and DL, respectively. For SMS MB=2 acquisitions, the scores were 3.7 [2.5, 4.5] and 3.3 [2.4, 4] for SMS-L1-SENSE and DL, respectively. DL results were inferior to CS reconstruction (p< 0.05). The main limitation in the image quality of the DL reconstruction was mild temporal flickering which may be improved with further optimization of the k-t sampling pattern or reconstruction network. All but 1 DL cases had clinically acceptable image quality (score > 3).

Figure 2 and Figure 3 present example cases from healthy volunteers with (SMS)-L1-SENSE and DL reconstruction. Good image quality was demonstrated.

The reconstruction time for DL was ~7 s per slice on an A100 GPU, while the (SMS-)L1-SENSE with 30 iterations on an Intel Xeon CPU (2.40 GHz) took ~45 minutes.

Conclusion:

The proposed image reconstruction network enabled rapid and high-quality image reconstruction for both SS and SMS MB=2 high-resolution Cartesian first-pass perfusion imaging at 3 T.