Synopsis

Keywords: Cancer, Image Reconstruction, Liver, Pancreas

Stack of Star acquisition is one of the most frequently used non-Cartesian k-space sampling methods due to its fast speed and robustness to motion. To further reduce the scan time or increase the temporal resolution, Stack of Star is often down-sampled with fewer spokes and advanced sampling patterns, such as golden angle acquisition. However, this makes Stack of Star prone to noise and streak artifacts, limiting the in-plane resolution and degrading diagnostic quality. In this work, we evaluated a deep-learning based stack-of-star method for free-breathing abdominal imaging and it shows improved diagnostic quality.

Introduction

Radial sampling is one of the most frequently used non-Cartesian k-space sampling methods due to its fast speed and robustness to motion. One of the 3D extensions of radial sampling is Stack of Star acquisition, which is often used in 3D breath-hold/free-breathing imaging and dynamic contrast enhanced imaging. (1-4)
To further reduce the scan time or increase the temporal resolution, Stack of Star is often down-sampled with fewer spokes and advanced sampling patterns, such as golden angle acquisition. However, this makes Stack of Star prone to noise and streak artifacts. Many investigators have employed sophisticated reconstruction techniques to produce artifact-free images, such as using compressed sensing (CS). However, CS reconstruction is computationally intensive, and the optimization of reconstruction parameters is also challenging.
In addition to the under-sampled k-space, respiratory motions and other non-rigid motions in the abdomen also cause streak artifacts. These streak artifacts are more obvious in post-contrast Stack of Star images due to the high signal intensity in moving tissues, reducing the lesion conspicuity. Truncation artifacts are another common artifact in Stack of Star images due to the limited acquisition matrix size and scan times. Low-pass filters are also applied in the conventional reconstruction method to minimize the truncation artifacts, but at the cost of in-plane resolution. These artifacts limit the in-plane resolution and lesion conspicuity, resulting in degraded diagnostic quality.
In this work, we evaluated a deep-learning based reconstruction method (DL Star) to improve the stack-of-star image quality by removing streak artifacts, truncation artifacts and noise in MR images.

Methods

Pre- and post-contrast liver and pancreas images were acquired in 5 patients using free-breathing 3D LAVA Star on a GE 3T MRI scanner (SIGNA Architect, GE Healthcare, Waukesha, WI) with IRB approval and written informed consent. The images were acquired with the following parameters, Matrix Size = 320 x320, Slice Thickness = 3.4 mm, Flip Angle = 12o, Repetition Time/Echo Time = 4 ms/1.92 ms, Bandwidth = 325.5 Hz/px, and Number of Averages = 0.7. A deep-learning based network (DL Star) was trained from a database of over 10,000 images to reconstruct images with a high signal-to-noise ratio, high spatial resolution, and reduced streak artifacts. A tunable noise reduction factor was offered to accommodate user preference. The DL Star network was embedded into the conventional reconstruction pathway to generate two sets of image series (conventional reconstructed and DL reconstructed images) from a single set of MR raw data.

Results and Discussions

Respiratory motion and other non-rigid motion in the abdomen often cause streak artifacts in the conventional stack-of-star images, especially in post-contrast images, as shown in Figure 1a. These streak artifacts reduce the conspicuity of lesions and anatomical details, degrading the diagnostic quality. To reduce these streak artifacts, DL Star reconstruction estimated and removed streak artifacts from the DL Star images, resulting in improved image quality and conspicuity of anatomical details, as shown in Figure 1b.
Truncation artifacts are often visible in stack-of-star images (Figure 2a) due to limited matrix size and scan times. Truncation artifacts are often suppressed with low-pass filters, further reducing the in-plane resolution. In contrast, DL Star reconstruction could suppress truncation artifacts without using low-pass filters, improving the in-plane resolution, as shown in Figure 2b.
Besides removing streak and truncation artifacts, DL Star could also reduce noise and improve the SNR, as shown in Figure 3. At low-field MRI, pre-contrast Stack of Star image could suffer from low SNR, requiring the acquisition of more spokes and longer scan times. Since DL Star could reduce noise, it has the potential to be used to reduce the scan time of Stack of Star at low-field MRI.
With the removal of streaks, truncation artifacts and noise, DL Star improved in-plane resolution and the conspicuity of lesions and anatomical details compared to the conventional reconstruction method. As shown in Figure 4, DL Star improved the visualization of pancreas with higher in-plane resolution and fewer streak and truncation artifacts. As shown in Figure 5, lesions also showed better conspicuity in DL Star post-contrast liver images due to efficient streak artifact removal.
The improvement is not limited to LAVA Star images, DL Star could also be used in other Stack of Star applications, such as DISCO Star for dynamic contrast enhancement imaging. In the future, we will evaluate DL Star with more patients.

Conclusion

The in-plane resolution and lesion conspicuity of 3D Stack of Star images are well improved with DL Star due to the promising noise, truncation, and streak artifacts reduction. DL Star showed the potential to improve the diagnostic quality of Stack of Star applications.

Acknowledgements

No acknowledgement found.