Echocardiography is a pivotal imaging tool for emergency medicine. Unfortunately, it suffers from poor image quality due to the intrinsic limitations of sonography systems. Towards this end, a better quality can be achieved at the cost of reduced frame rate by increasing the number of transmit/receive events and utilizing computationally expensive noise suppression algorithms. However, this visual quality and temporal resolution trade-off is a bottleneck for many echocardiography applications. Conventional acceleration methods, such as multi-line acquisition (MLA), work only for limited acceleration factors and produce blocking artifacts at a high frame rate. Accordingly, various machine learning algorithms have been designed to reduce blocking artifacts in MLA. These algorithms require access to either high-quality raw RF data or time-delayed baseband IQ data. Unfortunately, in many lower-end commercial systems, such data are not accessible. On the other hand, ultrasound images are badly affected by speckle noises which significantly reduces the image quality. We propose an image domain unsupervised deep learning framework using cycleGAN architecture for high quality accelerated echocardiography that simultaneously reduces the blocking artifacts and the speckle noise. The method is evaluated on real in-vivo and phantom data and achieves notable performance gain.

1. Unsupervised Deep Learning for Accelerated High
Quality Echocardiography
ABSTRACT
 Visual quality and temporal resolution trade-off is a bottleneck for many echocardiography applications.
 Conventional acceleration methods doesn’t remove noise and work only for limited acceleration factors.
 Various machine learning algorithms have been designed to reduce blocking artifacts using either high-quality raw RF data or time-delayed baseband IQ
data. Unfortunately, in many lower-end commercial systems, such data are not accessible.
 We propose an image domain framework using CycleGAN for high quality accelerated echocardiography that simultaneously reduces the blocking
artifacts and the speckle noise.
 The method is evaluated on real in-vivo and phantom data and achieves notable performance gain.
CONCLUSIONS
 A Universal image-domain-based method is proposed for the removal of speckle-noise and blocking artifacts from cardiac images.
 By synergistically learning from multiple datasets, the proposed method efficaciously generates a better quality image from as little as only 24 transmit events,
which is not possible from an analytic form of a standard MLA method.
 The design was evaluated through extensive tests on in vivo and phantom data for standard quality measures.
 In the current study, we intended to show a proof of concept for unpaired learning for high-quality ECHO and we believe that a clinical verification is necessary.
 The proposed scheme may substantially help in designing low-powered, high quality accelerated echocardiography systems.
ACKNOWLEDGEMENTS
This work was supported by the National Research
Foundation of Korea grant NRF-
2020R1A2B5B03001980.
REFERENCES
[1] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros, “Unpaired image-to-image
translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE international
conference on computer vision, 2017, pp. 2223–2232.
[2] Lei Zhu, Chi-Wing Fu, Michael S Brown, and Pheng-Ann Heng, “A non-local low-rank
framework for ultrasound speckle reduction,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, 2017, pp. 5650–5658.
Shujaat Khan1, Jaeyoung Huh1 and Jong Chul Ye1
1 Bio Imaging. Signal Processing & Learning Lab (BISPL), Dept. of Bio & Brain Engineering, KAIST, 291 Daehak-ro, Yuseong-gu, Daejeon
305-701, Korea.
https://bispl.weebly.com/
METHOD The architecture of the proposed network
single-line acquisition (SLA) mode with 96-tranmit events
(scanlines) and 64-receive and 128-transmit channels data.
Non-local low-rank based speckle free ultrasound [2]
Target styles
For training purpose, 55 phantom and 297 in vivo samples of 5 volunteers were used.
For the test purpose, 50 phantom and 192 in vivo frames of 2 volunteers were used.
Low-quality accelerated ECHO images are generated using 2, 3, 4 and 6-MLA.
Each US image has depth ranges between 2580 mm, and is scaled to 40 dB dynamic range.
In total, there are 55 + 297 × 5 = 1760 training and 50 + 192 × 5 = 1210 test images.
It is important to specify that unpaired dataset is used during the training.
RESULTS
Reconstruction results from unseen samples acquired from the ATS-539 phantom, and Cardiac region of a subject using SP1-5 phased
array probe operating at 3MHz center frequency on E-CUBE 12R US system (Alpinion, Korea).
The proposed method effectively enhances the visual quality for all acceleration factors.
 However, with higher acceleration factors, e.g., 6-MLA, the block artifacts are becoming prominent and
recovery to target quality is not ideal.
The average reconstruction time for an image was around 19:4 (milliseconds).
0 dB
-40 dB
6-MLA
1
0
0.5
Lateral length(mm)
0 70
35 105 140
Axial
depth(mm)
60
45
30
15
0
75
Input
2-MLA
Output
3-MLA 4-MLA
|Target-Input|
|Target-Output|
1
0
0.5
Lateral length(mm)
0 70
35 105 140
Axial
depth(mm)
60
45
30
15
0
75
Input
Output
|Target-Input|
|Target-Output|
(a)
B-mode
images
(b)
Normalized
residual
images
Target
(Filtered-SLA)
Target
(Filtered-SLA)
THEORY
Unsupervised learning using CycleGAN [1]
Here 𝜀𝜀𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐, and 𝜀𝜀𝐺𝐺𝐺𝐺𝐺𝐺 are cycle, and adversarial-loss, respectively, and 𝛼𝛼 is a loss mixing parameter.
Geometric view of unsupervised learning:
Here, the goal of optimal transport is to find 𝐺𝐺𝐴𝐴𝐴𝐴 and 𝐺𝐺𝐵𝐵𝐵𝐵 that minimize the
Wasserstein-1 distance between 𝜇𝜇 and 𝜇𝜇𝐵𝐵, and 𝜈𝜈 and 𝜈𝜈𝐴𝐴, respectively.
The objective is to learn optimal path for A B
𝐺𝐺𝐴𝐴𝐴𝐴
𝜈𝜈𝐴𝐴
𝜇𝜇𝐵𝐵
𝜇𝜇
𝜈𝜈
𝐺𝐺𝐵𝐵𝐵𝐵
A B

2. ABSTRACT
Visual quality and temporal resolution trade-off is a bottleneck for many echocardiography applications.
Conventional acceleration methods doesn’t remove noise and work only for limited acceleration factors.
Various machine learning algorithms have been designed to reduce blocking artifacts using either high-quality raw RF data or time-delayed baseband IQ data.
Unfortunately, in many lower-end commercial systems, such data are not accessible.
We propose an image domain framework using CycleGAN for high quality accelerated echocardiography that simultaneously reduces the blocking artifacts and
the speckle noise.
The method is evaluated on real in-vivo and phantom data and achieves notable performance gain.

3. THEORY Unsupervised learning using CycleGAN [1]
Here 𝜀𝜀𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐, and 𝜀𝜀𝐺𝐺𝐺𝐺𝐺𝐺 are cycle, and adversarial-loss, respectively, and 𝛼𝛼 is a loss mixing parameter.
Geometric view of unsupervised learning:
Here, the goal of optimal transport is to find 𝐺𝐺𝐴𝐴𝐴𝐴 and 𝐺𝐺𝐵𝐵𝐵𝐵 that
minimize the Wasserstein-1 distance between 𝜇𝜇 and 𝜇𝜇𝐵𝐵, and 𝜈𝜈
and 𝜈𝜈𝐴𝐴, respectively.
The objective is to learn optimal path for A B
𝐺𝐺𝐴𝐴𝐴𝐴
𝜈𝜈𝐴𝐴
𝜇𝜇𝐵𝐵
𝜇𝜇
𝜈𝜈
𝐺𝐺𝐵𝐵𝐵𝐵
A B

4. METHOD The architecture of the proposed network
single-line acquisition (SLA) mode with 96-tranmit events
(scanlines) and 64-receive and 128-transmit channels data.
Non-local low-rank based speckle free ultrasound [2]
Target styles
For training purpose, 55 phantom and 297 in vivo samples of 5 volunteers were used.
For the test purpose, 50 phantom and 192 in vivo frames of 2 volunteers were used.
Low-quality accelerated ECHO images are generated using 2, 3, 4 and 6-MLA.
Each US image has depth ranges between 2580 mm, and is scaled to 40 dB dynamic range.
In total, there are 55 + 297 × 5 = 1760 training and 50 + 192 × 5 = 1210 test images.
It is important to specify that unpaired dataset is used during the training.

5. RESULTS
Reconstruction results from unseen samples acquired from the ATS-539 phantom, and Cardiac region of a subject using SP1-5 phased array probe
operating at 3MHz center frequency on E-CUBE 12R US system (Alpinion, Korea).
The proposed method effectively enhances the visual quality for all acceleration factors.
 However, with higher acceleration factors, e.g., 6-MLA, the block artifacts are becoming prominent and recovery to target
quality is not ideal.
The average reconstruction time for an image was around 19:4 (milliseconds).
0 dB
-40 dB
6-MLA
1
0
0.5
Lateral length(mm)
0 70
35 105 140
Axial
depth(mm)
60
45
30
15
0
75
Input
2-MLA
Output
3-MLA 4-MLA
|Target-Input|
|Target-Output|
1
0
0.5
Lateral length(mm)
0 70
35 105 140
Axial
depth(mm)
60
45
30
15
0
75
Input
Output
|Target-Input|
|Target-Output|
(a)
B-mode
images
(b)
Normalized
residual
images
Target
(Filtered-SLA)
Target
(Filtered-SLA)

6. CONCLUSIONS
A Universal image-domain-based method is proposed for the removal of speckle-noise and blocking artifacts from cardiac images.
By synergistically learning from multiple datasets, the proposed method efficaciously generates a better quality image from as little as only 24 transmit events,
which is not possible from an analytic form of a standard MLA method.
The design was evaluated through extensive tests on in vivo and phantom data for standard quality measures.
In the current study, we intended to show a proof of concept for unpaired learning for high-quality ECHO and we believe that a clinical verification is necessary.
The proposed scheme may substantially help in designing low-powered, high quality accelerated echocardiography systems.
ACKNOWLEDGEMENTS
This work was supported by the National Research
Foundation of Korea grant NRF-2020R1A2B5B03001980.
REFERENCES
[1] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros, “Unpaired image-to-image
translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE international
conference on computer vision, 2017, pp. 2223–2232.
[2] Lei Zhu, Chi-Wing Fu, Michael S Brown, and Pheng-Ann Heng, “A non-local low-rank
framework for ultrasound speckle reduction,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, 2017, pp. 5650–5658.