Introduction

Single-shot echo planar diffusion weighted imaging sequence (EPI-DWI) is often acquired with small matrix size to shorten the scan time and/or EPI readout duration. The truncated matrix aggravates Gibbs ringing artifact. Typically, low pass filters in k-space and/or image space are used to reduce the ringing artifact, at the expense of image blurring. In a previous study[1, 2, 3], a high-resolution deep learning reconstruction (HR-DLR) was proposed, combining ZIP with two CNN’s: the first one trained to reduce image noise, the second one to reduce Gibbs artifacts. The HR-DLR was evaluated on knee images using 2D Fast Spin Echo sequence[1, 2] and brain with EPI-DWI[3]. Those studies showed better SNR and sharpness with reduced Gibbs artifacts versus standard reconstructions used in routine clinical practice. The goal of this work was to test that HR-DLR method enables high-resolution imaging or shorted acquisition time preserving structural detail and reducing Gibbs artifact on brain EPI-DWI.

Method

The reconstruction pipeline of the proposed method is illustrated in Figure 1, same as the previous studies[1, 2, 3]. Although the two neural networks were developed separately, they shared the same ground-truth dataset for training. This ground-truth dataset was collected with higher SNR and resolution than typical clinical images, by averaging repeated scans. The acquisitions included multiple body parts (brain, spine, shoulder, elbow, wrist, hip, knee, and ankle), contrast weightings (DWI, T1w, T2w, FLAIR, T2*w, PDw, etc.), and field strengths (3T and 1.5T). HR-DLR was tested on b0 and isotropic diffusion images(isoDWI) with clinical standard protocol and fast protocol.
The brain of a healthy volunteer was scanned at 1.5T (Vantage Orian XGO, Canon Medical Systems Corporation, Tochigi, Japan) with a 16ch head coil. The study was approved by our institutional review board and informed consent was obtained. The relevant parameters of the standard protocol are TR/TE=5700/75[ms], echo train spacing (ETS)=0.9[ms], Bandwidth=1302[Hz], b-value=0 and 1000 in 3-axis, acquisition matrix=160, and number of acquisition (NAQ)=3, while those of the fast protocol are TR/TE=4603/75[ms], ETS=0.7[ms], Bandwidth=1953[Hz], b-value=0 and 1000 in 3-axis, acquisition matrix =128, and NAQ=1. The fast protocol scan time is 42[sec], while the standard protocol is 92[sec].
Three metrics were measured:
1. Edge Slope Width(ESW)[4] in line ROIs: $$(SI_{max}–SI_{min})/Slope,$$ where $$$Slope$$$ is a rate of change of signal intensity near the center of edge, and $$$SI_{max}$$$ and $$$SI_{max}$$$ are the average signal intensity of three neighboring points in the plateaus before and after the edge. Four ROIs were drawn on edge of lateral ventricle. Basically, a smaller ESW means the edge is sharp, i.e., it suddenly rises or falls.
2. Ringing Variable Mean(RVM)[5] in line ROIs which are the mean absolute value of differential of the profile curves: $$\frac{1}{N-1}\sum_{i=1}^{N-1}\left|\frac{SI[i+1]-SI[i]}{\Delta{x}}\right|,$$ where $$$SI[i]$$$ is the signal intensity at the $$$i$$$-th point. Four ROIs were drawn on regions demonstrating Gibbs artifacts in images without filtering (NONE). Generally, a smaller RVM means less ringing artifacts.
3. Signal-to-Noise (SNR) in circle ROIs which are the mean signal intensity divided by the standard deviation:
Two ROIs were drawn in the thalamus.
Test#1. HR-DLR for High Matrix Size:
The three metrics are measured on the standard protocol and compared between ZIP method using an upscaling factor (in both in-plane directions) of 2 without any filter (NONEx2), ZIPx2 with a low pass filter (LPFx2), and HR-DLR with upscaling factor of 3 (HR-DLRx3).
Test#2. HR-DLR for Short Time Acquisition:
The three metrics are measured and compared between LPFx2 of the standard protocol(Standard LPFx2), NONEx2 of the fast protocol(Fast NONEx2) and HR-DLRx3 of the fast protocol(Fast HR-DLRx3).

Results

The resulting images from the retrospective test are shown in Figure 4 and 5. In the Test #1, HR-DLRx3 showed better sharpness(smaller ESW) and SNR with reducing Gibbs artifact(smaller RVM) compared to both conventional methods of LPFx2 and NONEx2.In the Test #2, Fast HR-DLRx3 achieved better or equivalent image quality to Standard LPFx2. Fast HR-DLRx3 reduced Gibbs artifact as well as LPFx2 compared to Fast NONEx2 in terms of the RVM and image appearance. Fast HR-DLR showed similar sharpness with LPFx2 in terms of both ESW and appearance. The Fast HR-DLR achieved better SNR compared to both LPFx2 and Fast NONEx2.

Conclusion

This work shows that HR-DLR improves sharpness of EPI-DWI with less or similar Gibbs ringing artifacts compared to conventional ZIP with Low pass filtered images via the metrics of ESW and RVM. It also shows HR-DLR enables shortenerd acquisition times by lowering matrix size and NAQ, while upsampling allows reconstruction of high-resolution images, preserving structural detail and removing the noise.

Acknowledgements

No acknowledgement found.