Hiroshi Kusahara^{1}, Yuki Takai^{1}, Kensuke Shinoda^{1}, and Yoshimori Kassai^{1}

In this study to the authors adapted the variable-TE cDWI(vTE-cDWI) technique applying denoise approach with deep learning reconstruction(dDLR) to the abdominal region, using ADC-map, T2-map and T1-map with IR-based images. The algorithm under evaluation allows computing diffusion images for arbitrary combinations of TE, b-value and TI based on four acquisitions(4-points method). This technique was shown to generate vTE-cDWI with higher SNR compared to the acquired DWI, and dDLR increased the SNR more, as well as obtain ADC-maps and T1-maps with optimal TI for any arbitrary tissue. The clinical benefits of the method and results on volunteers are discussed.

A computed diffusion weighted imaging(cDWI) was previously presented, demonstrating improvement in lesion
detection_{[1-3]}. This algorithm works by computing high
b-value-equivalent(*b*) images from relatively
low-b images. Recently a variable-TE cDWI(vTE-cDWI) technique was
proposed for the brain allowing the computation of DWIs with arbitrary combinations
of TE and b_{[4]}. This
method can control the T2-shine-through effects and optimize the SNR
and CNR of tissues. However,
for creating higher b value images with higher accuracy, the original image
with higher SNR is necessary.

Noise reduction technique is one of the methods
to solve this problem. We have developed the denoise approach
with deep learning reconstruction(dDLR), as
noise reduction technique based on deep learning, using high SNR images and
images with various intensities of noise added_{[5][6]}.

In this study we extended the vTE-cDWI
technique, applying dDLR to each map images(T1,T2,ADC), adapting it to the
abdomen_{[7]}. Improved CNR and SNR is achieved by computing images with
arbitrary TE and b, while reducing the T2-shine-through effects and allowing for
selective tissue suppression through the arbitrary TI.

*Theory:*
For
a SEEPI2D(mDWI) with set TE and b, the intensity at voxel location
characterized by known T2 and ADC parameters is described by:

S(TE,b)=k⋅exp((−TE)/T2)⋅exp(−b⋅ADC)⋅⋅⋅(1)

k being a constant for the voxel. Similarly, IR-based mDWI signal intensity at given TI, TE, and b when TR is set to infinity is described by:

S(TI,TE,b)=k⋅(1−2exp(−TI/T1))⋅exp(−TE/T2)⋅exp(−b⋅ADC)⋅⋅⋅(2)

A 4-point cDWI method was applied, along the algorithm shown in Fig.1, as follows:

*a)Same-TE cDWI _{[1-3]}: *

A 2-points technique is applied, measuring mDWI signals
at the same TE=TE2 for two different
b-values (*b1<b2*), and solving (1) for the ADC:

ADC=−ln[S2(TE2,b2)/S1(TE2,b1)]/(b2−b1).⋅⋅⋅(3)

The cDWI image at the original TE=TE2 and for arbitrary b=*bc* is obtained by replacing (3)
into (1):

Sc(TE2,bc)=S1(TE2,b1)⋅exp[−(bc−b1)⋅ADC].⋅⋅⋅(4)

*B)Different-TE cDWI (vTE-cDWI) _{[4]}:*

The tissue T2 can be obtained from (1) by measuring two mDWI with the same b=b1 and different TEs (TE1<TE2):

1/T2=−ln[S1(TE2,b1)/S3(TE1,b1)]/(TE2−TE1).⋅⋅⋅(5)

The diffusion image at
arbitrary TE=*TEc* and b=*bc* is obtained replacing the calculated
ADC (3) and T2 (5) in (1):

Sc(TEc,bc)=S3(TE1,b1)⋅exp[−(TEc−TE1)/T2]⋅exp[−(bc−b1)⋅ADC].⋅⋅⋅(6)

(6) allows to compute sTE-cDWI images free of T2-effects(*TEc*=0).

*C)IR-based STIR vTE-cDWI:
*

T1 is derived
from (2) with the IR term in *S4(TI1,TE1,b1)*
by setting TI=TI1, the SE signal of PDW *S3(TE1,b2)*, and selecting
a sufficiently long* TR _{2}*:

1/T1=(−ln[(1−S4/S3)/2])/TI1⋅⋅⋅(7)

By replacing the calculated ADC (3), T2 (5), and T1 (7)
in (2), we can compute the diffusion image at arbitrary TI=*TIc*, TE=*TEc* and b=*bc* as:

Sc(TIc,TEc,bc)=S3⋅[1−2⋅exp(−TIc/T1)]⋅exp[−(TEc−TE1)/T2]⋅exp[−(bc−b1)⋅ADC].⋅⋅⋅(8)

*Volunteers study:* The models described above were validated on three volunteers. Abdominal axial scans through the liver were acquired
on a Canon 3T MR system using a single-shot SEEPI2D sequence. Acquisition
parameters for the 4-points mDWI scans were: TE1=25ms, b1=0 without inversion
and TE1=25ms, b1=0 with TI1=600ms, TE2=39ms with b1=0 and b2=800s/mm^{2}.
TE3=45ms and TE4=60ms with b1=0 were also acquired for a more accurate
assessment of the T2 map. The T1 map was calculated with TI-off(=infinity) and
TI=600ms for the same TE1=25ms. The cDWI and STIR-cDWI signals for *bc*=800, and 1500s/mm^{2} were
calculated at different *TEc*. In
addition, dDLR was applied to each map, and same cDWI and STIR-cDWI images were calculated
at the image after adaptation.
ROIs were drawn on the liver, cholecyst, kidney, and
the erector spinae muscle. SNR related to kidney-SD and CNR with respect to the
muscle were measured for each ROI(Table.1).

Volunteers’ scans demonstrated an increased liver-to-muscle CNR for shorter TEs, suggesting that the cDWI can improve the characterization of tissues with short T2. SNRs in dDLR image for vTE-cDWI increased at all condition compared to mDWI and vTE-cDWI without dDLR(Fig.2). The shoter-TE cDWI with dDLR has a PDwi contrast, therefore a decreased signal for (TE, b)=(25, 800) is descriptive of the diffusion properties with higher SNR and less T2 shine-thorough effects.

SNRs in the liver and cholecyst for the IR-based STIR vTE-cDWI(method-C) increased at (TI1,TE,b)=(280,60,1500) compared to the STIR-mDWI. SNRs in dDLR image for STIR vTE-cDWI also increased at all condition compared to without dDLR. The IR-based STIR vTE-cDWI with dDLR in the abdomen is meant for fat suppression or background suppression, and benefits from controlling TE, b and TI(Fig.3,Fig4). This technique demonstrated an improved CNR compared to acquired images because of the contribution of the T1 map.

[1] *Blackledge
MD, Leach MO, Collins DJ, et al. Computed Diffusion-weighted MR Imaging May
Improve Tumor Detection. Radiology: 261,573-581(2011).
*

[2]
*Ueno Y, Takahashi S, Kitajima K, Kimura T
et al. Computed diffusion-weighted imaging using 3-T magnetic resonance imaging
for prostate cancer diagnosis. Eur Radiol 23,3509-16 (2013).*

[3] *Kimura T, Machi Y.
Computed Diffusion Weighted Imaging Under Rician Noise Distribution. In:
Proceedings of the 20th Annual Meeting of ISMRM, 2012;p3574.
*

[4]
*Kimura T, Machi Y et al. A short-TE Computed Diffusion Imaging
(cDWI). In: Proceedings of the 23th Annual Meeting of ISMRM, 2015;p2929.
*

[5] *K. Isogawa, T. Ida, T.
Shiodera and T. Takeguchi, "Deep Shrinkage Convolutional Neural Network
for Adaptive Noise Reduction," IEEE Signal Processing Letters, vol. 25,
no. 2, pp. 224-228, 2018.
*

[6]
*K. Isogawa, T. Ida, T. Shiodera et al. Deep Convolutional Neural Network for Image Denoising. In:
Proceedings of the 26th Annual Meeting of ISMRM, 2018;823.
*

[7] *Warntjes JBM et al. Rapid Magnetic Resonance
Quantification on the Brain. Magn Reson Med 60:320–329 (2008).*

Fig.1 vTE-cDWI and STIR-cDWI calculation flow for 4-point (maximum)
method

Table.1 ROIs were drawn on the liver, cholecyst, kidney, and the erector
spinae muscle. SNR related to kidney-SD and CNR with respect to the muscle were
measured for each ROI.

Fig.2 Comparison between acquired and computed diffusion images
at same TE and b value with/without DLR.

Fig.3 Comparison between acquired and computed STIR diffusion images
with matching scan parameters with/without DLR.

Fig.4 Comparison of background suppression on CO-MPR between acquired
and computed STIR diffusion images with matching scan parameters and DLR with
computed diffusion image.