Chinese Journal of Liquid Crystals and Displays, Volume. 37, Issue 12, 1626(2022)
Diabetic retinopathy classification method based on cost sensitive regularization and EfficientNet
Fig. 3. Different levels of DR images.（a）Normal without obvious lesions；（b）Mild non-proliferative DR；（c）Moderate non-proliferative DR；（d）Severe non-proliferative DR；（e）Proliferative DR.
Fig. 4. Diabetic retinopathy classification model of improved EfficientNet
Fig. 5. Performance of EfficientNet-CSR in the EyePACS test set.（a）ROC curves；（b）Confusion matrix.
Table 1. Data distribution of the Kaggle EyePACS
View tableView in ArticleTable 1. Data distribution of the Kaggle EyePACS
等级 名称 图片数 占比/% 0 No DR 25 810 73.48 1 Mild DR 2 443 6.95 2 Moderate DR 5 292 15.07 3 Severe DR 873 2.49 4 Proliferative DR 708 2.01
Table 2. EfficientNet network performance with cross entropy loss function and focal loss function
View tableView in ArticleTable 2. EfficientNet network performance with cross entropy loss function and focal loss function
损失函数 ACC kappa分数 AUC 交叉熵损失函数 0.88 0.85 0.91 焦点损失函数 0.90 0.86 0.93
Table 3. Comparison of experimental results
View tableView in ArticleTable 3. Comparison of experimental results
网络结构 ACC kappa分数 AUC Inception V3 0.82 0.78 0.86 Inception V3-CSR 0.84 0.82 0.88 EfficientNet 0.86 0.81 0.90 EfficientNet-CSR 0.90 0.86 0.93
Table 4. Performance of four Network models on two data sets
View tableView in ArticleTable 4. Performance of four Network models on two data sets
数据集 ResNeXt50 Inception V3 Lands等人结果［ 16 ］EfficientNet-CSR EyePACS 0.77/0.84 0.78/0.86 0.81/0.91 0.86/0.93 Messidor-2 0.80/0.85 0.79/0.88 0.83/0.92 0.88/0.95
Get Citation
Copy Citation Text
Ming-zhi WANG, Zhi-qiang MA, Feng-feng ZHAO, Yong-jie WANG, Ji-feng GUO. Diabetic retinopathy classification method based on cost sensitive regularization and EfficientNet[J]. Chinese Journal of Liquid Crystals and Displays, 2022, 37(12): 1626
Category: Research Articles
Received: May. 11, 2022
Accepted: --
Published Online: Nov. 30, 2022
The Author Email: Zhi-qiang MA (1791105996@qq.com)
