Fundus Image Classification Research Based on Ensemble Convolutional Neural Network and Vision Transformer

Yuan Yuan; Minghui Chen; Shuting Ke; Teng Wang; Longxi He; Linjie Lü; Hao Sun; Jiannan Liu

doi:10.3788/CJL202249.2007205

Chinese Journal of Lasers, Volume. 49, Issue 20, 2007205(2022)

Fundus Image Classification Research Based on Ensemble Convolutional Neural Network and Vision Transformer

Yuan Yuan, Minghui Chen^*, Shuting Ke, Teng Wang, Longxi He, Linjie Lü, Hao Sun, and Jiannan Liu

Shanghai Engineering Research Center of Interventional Medical, Ministry of Education of Medical Optical Engineering Center, School of Health Sciences and Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(10)

Fig. 1. Overview of Vit model

Download full size

Fig. 2. Structure of MBConv and Fused-MBConv. (a) MBConv; (b) Fused-MBConv

Download full size

Fig. 3. SimAM 3D attention weights

Download full size

Fig. 4. Confusion matrix of Vit and EfficientNetV2-S models. (a) Vit; (b) EfficientNetV2-S

Download full size

Fig. 5. Heatmap of abnormal fundus images in dataset. (a) Abnormalities in optic disc; (b) abnormalities in macular

Download full size

Table 1. EfficientNetV2-S architecture

View table

Table 1. EfficientNetV2-S architecture

Stage	Operator	Stride	Number of channels	Number of layers
0	Conv 3×3	2	24	1
1	Fused-MBConv1，k 3×3	1	24	2
2	Fused-MBConv4，k 3×3	2	48	4
3	Fused-MBConv4，k 3×3	2	64	4
4	MBConv4, k 3×3, SimAM	2	128	6
5	MBConv6, k 3×3, SimAM	1	160	9
6	MBConv6, k 3×3, SimAM	2	272	15
7	Conv 1×1 & Pooling & FC	－	1792	1

Table 2. Fundus dataset
View table
Table 2. Fundus dataset
Degree of illness Number of training images Number of testing images Total number of images
Normal 2818 409 3227
DR 1389 203 1592
ARMD 147 49 196
Myopia 234 35 269
Cataract 262 43 305

Table 3. Accuracy, precision, and specificity of Vit, EfficientNetV2-S, and EfficientNet-Vit models
View table
Table 3. Accuracy, precision, and specificity of Vit, EfficientNetV2-S, and EfficientNet-Vit models
Model Accuracy /% Precision /% Specificity /% Training time /h
Vit 91.1 86.4 97.2 11.0
EfficientNetV2-S 92.2 87.6 97.5 9.2
EfficientNet-Vit 92.7 88.3 98.1 －

Table 4. Comparison of accuracy indexes of different models
View table
Table 4. Comparison of accuracy indexes of different models
Model Accuracy /%
Resnet50 87.3
Densenet121 89.5
ResNeSt-101 90.7
EfficientNet-B0 91.3
TNT-B 91.1
EfficientNet-Vit 92.7

Table 5. Comparison of accuracy indexes of models with different weighted factors
View table
Table 5. Comparison of accuracy indexes of models with different weighted factors
Weighted factor Accuracy /%
0.3, 0.7 92.0
0.4, 0.6 92.7
0.5, 0.5 91.6

Tools

Get Citation

Copy Citation Text

Yuan Yuan, Minghui Chen, Shuting Ke, Teng Wang, Longxi He, Linjie Lü, Hao Sun, Jiannan Liu. Fundus Image Classification Research Based on Ensemble Convolutional Neural Network and Vision Transformer[J]. Chinese Journal of Lasers, 2022, 49(20): 2007205

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites