Deep Learning Feature Fusion-Based Retina Image Classification

Tianfu Zhang; Shuncong Zhong; Chaoming Lian; Ning Zhou; Maosong Xie

doi:10.3788/LOP57.241025

Laser & Optoelectronics Progress, Volume. 57, Issue 24, 241025(2020)

Deep Learning Feature Fusion-Based Retina Image Classification

Tianfu Zhang¹, Shuncong Zhong^1、*, Chaoming Lian¹, Ning Zhou¹, and Maosong Xie²

Author Affiliations

¹School of Mechanical Engineering and Automation, Fuzhou University, Fuzhou, Fujian 350108, China

²First Affiliated Hospital of Fujian Medical University, Fuzhou, Fujian 350000, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Convolutional neural network model

Download full size

Fig. 2. Convolutional layer structure comparison. (a) Ordinary convolution; (b) depth separable convolution

Download full size

Fig. 3. Sample graphs of the dataset. (a) CNV; (b) DME; (c) DRUSEN; (d) NORMAL

Download full size

Fig. 4. Image preprocessing. (a) Original image of OCT retina; (b)mean-shift removed speckle image

Download full size

Fig. 5. Validation result curves. (a) Validation accuracy curve; (b) validation loss curve

Download full size

Fig. 6. Confusion matrix. (a) Confusion matrix without weighted loss function; (b) confusion matrix with weighted loss function

Download full size

Fig. 7. Visual heat maps

Download full size

Fig. 8. Confusion matrix. (a) Confusion matrix for GAPNet model; (b) confusion matrix for RongheNet model

Download full size

Table 1. Division of OCT dataset
View table
Table 1. Division of OCT dataset
Image type Number of images
Train Validation Test Total
CNV 26043 7441 3721 37205
DME 7943 2270 1135 11348
DRUSEN 6031 1723 862 8616
NORMAL 18420 5263 2632 26315
Total 58437 16697 8350 83484

Table 2. Comparison of classification indexes of different training methods unit: %
View table
Table 2. Comparison of classification indexes of different training methods unit: %
Network type Category weight Accuracy MA Precision Recall
SlimNet No 89.9 82.5 88.0 82.0
SlimNet Yes 90.2 87.2 85.0 87.0
BnL2Net Yes 93.0 91.6 92.0 93.0
GAPNet Yes 94.9 93.3 92.0 93.0
RongheNet Yes 97.2 97.0 95.0 97.0

Table 3. Performance comparison of different models unit: %
View table
Table 3. Performance comparison of different models unit: %
Method MA Precision Recall
Wang et al^[6] 89.6 90.8 88.6
Bhowmik et al^[7] 92.6 94.0 94.0
Yu et al^[8] 94.1 94.6 90.6
Ours 97.0 95.0 97.0

Tools

Get Citation

Copy Citation Text

Tianfu Zhang, Shuncong Zhong, Chaoming Lian, Ning Zhou, Maosong Xie. Deep Learning Feature Fusion-Based Retina Image Classification[J]. Laser & Optoelectronics Progress, 2020, 57(24): 241025

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites