Super-Resolution Reconstruction of Magnetic Resonance Image Based on Deep Learning

Mengxue Pan; Ning Qu; Yeru Xia; Deyong Yang; Hongyu Wang; Wenlong Liu

doi:10.3788/LOP202259.2217001

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 22 >Page 2217001 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 22, 2217001 (2022)

Super-Resolution Reconstruction of Magnetic Resonance Image Based on Deep Learning

Mengxue Pan, Ning Qu, Yeru Xia, Deyong Yang, Hongyu Wang, and Wenlong Liu^*

Author Affiliations

Faculty of Electronic Information and Electrical Engineering, Dalian University of Technology, Dalian 116024, Liaoning, China

show less

DOI: 10.3788/LOP202259.2217001 Cite this Article Set citation alerts

Mengxue Pan, Ning Qu, Yeru Xia, Deyong Yang, Hongyu Wang, Wenlong Liu. Super-Resolution Reconstruction of Magnetic Resonance Image Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2217001 Copy Citation Text

show less

Fig. 1. Network structure

Download full size

Fig. 2. Convolution residual block and residual block. (a) Convolution residual block; (b) residual block

Download full size

Fig. 3. Sub-pixel convolution

Download full size

Fig. 4. Image 1 reconstruction results with different algorithms. (a) Original image; (b) Bicubic; (c) FSRCNN; (d) EDSR; (e) SRResNet; (f) proposed algorithm

Download full size

Fig. 5. Image 2 reconstruction results with different algorithms. (a) Original image; (b) Bicubic; (c) FSRCNN; (d) EDSR; (e) SRResNet; (f) proposed algorithm

Download full size

Fig. 6. Image 3 reconstruction results with different algorithms. (a) Original image; (b) Bicubic; (c) FSRCNN; (d) EDSR; (e) SRResNet; (f) proposed algorithm

Download full size

Layer	Input	Filter	Output
Conv-RB1	$128 \times 128 \times 64$	$3 \times 3 \times 64$	$128 \times 128 \times 64$
Conv-RB2	$128 \times 128 \times 64$	$3 \times 3 \times 128$	$128 \times 128 \times 128$
Conv-RB3	$128 \times 128 \times 128$	$3 \times 3 \times 256$	$128 \times 128 \times 256$
Conv-RB4	$128 \times 128 \times 256$	$3 \times 3 \times 512$	$128 \times 128 \times 512$
Conv-RB5	$128 \times 128 \times 512$	$3 \times 3 \times 256$	$128 \times 128 \times 256$
Conv-RB6	$128 \times 128 \times 256$	$3 \times 3 \times 128$	$128 \times 128 \times 128$
Conv-RB7	$128 \times 128 \times 128$	$3 \times 3 \times 64$	$128 \times 128 \times 64$

Table 1. Convolution residual block parameter setting

Grade	Absolute measurement scale	Detail	Score
1	Excellent	The best in the group	5
2	Good	Better than the average in the group	4
3	Average	Group average	3
４	Fair	Worse than the average in the group	2
５	Poor	Worse in the group	1

Table 2. Image subjective evaluation form

MR image	Bicubic	FSRCNN	EDSR	SRResNet	Proposed method
1	1.0	2.0	3.0	4.0	5.0
2	1.0	1.8	3.0	4.2	4.8
3	1.0	1.8	3.0	4.2	5.0

Table 3. Comparison of subjective evaluation values of super-resolution reconstruction methods

MR image	Bicubic	FSRCNN	EDSR	SRResNet	Proposed method
1	28.69	29.93	29.23	29.25	31.18
2	23.10	24.23	24.17	24.75	29.88
3	26.40	25.01	26.87	26.62	29.12

Table 4. Comparison of PSNR values of various super-resolution reconstruction

MR image	Bicubic	FSRCNN	EDSR	SRResNet	Proposed method
1	3931202	6710233	9213167	12092766	13343760
2	5533140	8946652	12582732	14374929	16688306
3	4547242	7907218	11300894	12340983	15936792

Table 5. Comparison of energy gradient values of various super-resolution reconstruction

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information