Unsupervised Denoising of Retinal OCT Images Based on Deep Learning

Guangyi Wu; Zhuoqun Yuan; Yanmei Liang

doi:10.3788/AOS230720

Acta Optica Sinica, Volume. 43, Issue 20, 2010002(2023)

Unsupervised Denoising of Retinal OCT Images Based on Deep Learning

Guangyi Wu, Zhuoqun Yuan, and Yanmei Liang^*

Institute of Modern Optics, Nankai University, Tianjin Key Laboratory of Micro-Scale Optical Information Science and Technology, Tianjin 300350, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Network structure of DRSA-Net

Download full size

Fig. 2. Flow chart of retinal OCT image unsupervised training experiment

Download full size

Fig. 3. Noise reduction in retinal OCT images. (a) Original noisy image; (b) denoised images of BM3D; (c) denoised image of DnCNN-N2N; (d) 5-frame average ground truth images; (e) denoised image of U-Net-N2N; (f) denoised image of Ours-N2N

Download full size

Fig. 4. Noise reduction results of supervised learning and unsupervised learning retinal OCT images. (a) Original noisy image; (b) denoised image of DnCNN-N2C; (c) denoised image of U-Net-N2C; (d) denoised image of Ours-N2C; (e) 5-frame average ground truth image; (f) denoised image of DnCNN-N2N; (g) denoised image of U-Net-N2N; (h) denoised image of Ours-N2N

Download full size

Fig. 5. Unsupervised learning generalization ability test. (a) Original noisy image; (b) denoised images of BM3D; (c) denoised image of DnCNN-N2N; (d) 40-frame average ground truth images; (e) denoised image of U-Net-N2N; (f) denoised image of Ours-N2N

Download full size

Fig. 6. Comparison of generalization ability between supervised and unsupervised learning. (a) Denoised image of DnCNN-N2C; (b) denoised image of U-Net-N2C; (c) denoised image of Ours-N2C; (d) denoised image of DnCNN-N2N; (e) denoised image of U-Net-N2N; (f) denoised image of Ours-N2N

Download full size

Table 1. Results of supervised learning and unsupervised learning denoising numerical evaluation

View table

Table 1. Results of supervised learning and unsupervised learning denoising numerical evaluation

Method		PSNR	SSIM	EPI	ENL	Time /s
Baseline	Noisy	19.073±0.065	0.391±0.002		14.938±1.693
Baseline	BM3D	23.937±0.181	0.253±0.101	0.281±0.051	434.819±111.647	127
Supervised learning model	DnCNN	25.418±0.213	0.506±0.007	0.295±0.023	220.663±14.472	0.48
	U-Net	24.852±0.322	0.498±0.006	0.304±0.024	255.427±20.642	0.70
	Ours	25.447±0.191	0.494±0.007	0.312±0.025	279.760±27.946	0.53
Unsupervised N2N model	DnCNN	24.234±0.183	0.284±0.007	0.242±0.027	768.128±137.606	0.48
	U-Net	24.543±0.212	0.287±0.008	0.243±0.028	1601.956±573.328	0.70
	Ours	24.582±0.225	0.289±0.008	0.262±0.026	1304.384±466.983	0.53

Table 2. Results of supervised learning and unsupervised learning denoising numerical evaluation

View table

Table 2. Results of supervised learning and unsupervised learning denoising numerical evaluation

Method		PSNR	SSIM	EPI	ENL	Time /s
Baseline	Noisy	18.193±0.273	0.134±0.024		14.074±3.681
Baseline	BM3D	28.035±1.457	0.550±0.044	0.268±0.035	418.037±106.729	256
Supervised learning model	DnCNN	28.699±1.301	0.554±0.036	0.277±0.027	216.478±37.898	0.79
	U-Net	27.956±1.186	0.569±0.035	0.287±0.028	241.361±31.499	1.04
	Ours	29.317±1.419	0.593±0.034	0.293±0.030	262.771±34.762	0.92
Unsupervised N2N model	DnCNN	29.673±1.195	0.665±0.012	0.166±0.037	628.151±182.529	0.79
	U-Net	30.950±1.565	0.702±0.012	0.176±0.036	1266.897±338.543	1.04
	Ours	31.172±1.706	0.706±0.013	0.194±0.035	1029.639±220.714	0.92

Table 3. Ablation experimental results of different modules of network

View table

Table 3. Ablation experimental results of different modules of network

Module	PSNR	SSIM	EPI	ENL	Time/s
DEB+GAB+RB	28.962±1.632	0.703±0.015	0.131±0.031	159.334±13.502	0.73
LSAB+GAB+RB	28.737±1.264	0.613±0.026	0.133±0.072	879.870±207.982	0.69
LSAB+DEB+RB	30.301±1.446	0.647±0.012	0.150±0.035	903.572±343.217	0.73
LSAB+DEB+GAB	30.802±1.427	0.701±0.012	0.127±0.033	937.421±225.073	0.75
LSAB+DEB+GAB+RB	31.172±1.706	0.706±0.013	0.194±0.035	1029.639±220.714	0.92

Tools

Get Citation

Copy Citation Text

Guangyi Wu, Zhuoqun Yuan, Yanmei Liang. Unsupervised Denoising of Retinal OCT Images Based on Deep Learning[J]. Acta Optica Sinica, 2023, 43(20): 2010002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Image Processing

Received: Mar. 27, 2023

Accepted: May. 6, 2023

Published Online: Oct. 23, 2023

The Author Email: Yanmei Liang (ymliang@nankai.edu.cn)

DOI:10.3788/AOS230720

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology