EfficientNet in diabetic retinopathy detection: impact of pre-training scale
Introduction: Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients worldwide. Timely and accurate detection of referable diabetic retinopathy (RDR) is essential for preventing blindness. In this context, deep learning technologies offer promising advancements in the autom...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English English |
| Published: |
Universiti Putra Malaysia Press
2024
|
| Subjects: | |
| Online Access: | http://irep.iium.edu.my/117717/7/117717_EfficientNet%20in%20diabetic%20retinopathy%20detection.pdf http://irep.iium.edu.my/117717/13/117717_EfficientNet%20in%20diabetic%20retinopathy%20detection_Scopus.pdf http://irep.iium.edu.my/117717/ https://medic.upm.edu.my/upload/dokumen/2024123014280818_MJMHS_0148.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Introduction: Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients worldwide. Timely and accurate detection of referable diabetic retinopathy (RDR) is essential for preventing blindness. In this context, deep learning technologies offer promising advancements in the automated detection of DR. This study explores the effect of pre-training dataset sizes on the diagnostic accuracy of the EfficientNet architecture for RDR identification. Methods: We utilised the EyePACS dataset containing 23,252 retinal images, including 4,343 RDR instances, for training our models. These models were then tested on the APTOS dataset, comprising 3,662 images with 1,487 cases of RDR. Two variants of EfficientNet, one pre-trained on an ImageNet-1K dataset and the other on an ImageNet-21K dataset, were compared based on their sensitivity, specificity, and AUC. Results: The EfficientNet variant pretrained on the ImageNet-1K dataset achieved a sensitivity of 97.71%, specificity of 83.13%, and an AUC of 0.901. In comparison, the variant pre-trained on the ImageNet-21K dataset demonstrated a slightly improved sensitivity of 98.79%, but a reduced specificity of 80.83%, with a comparable AUC of 0.898. Conclusion: Our study demonstrates that deep learning is a valuable tool for RDR detection, with pre-training on larger datasets resulting in modest improvements in sensitivity. However, the difference in pre-training dataset size did not substantially alter the AUC, indicating that additional factors may contribute to the overall effectiveness of the models. These results emphasize the potential of deep learning in enhancing DR screening and diagnosis, with implications for reducing the burden of diabetes-related vision loss. Malaysian Journal of Medicine and Health Sciences (2024) 20(SUPP10): 133-138. doi:10.47836/mjmhs.20. s10.18 |
|---|