EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

EfficientNet and Swin Transformer Fusion for Diabetic Retinopathy Screening

B. krishna Sahithi, Achalla Sruthi Chandrika, Manju Venugopalan

Procedia Computer Science · 2026 · 283 · 4560-4569

DOI: 10.1016/j.procs.2026.06.512 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Procedia Computer Science, 2026. Verfasst von B. krishna Sahithi und 2 weiteren Autor:innen. Die Arbeit nennt 22 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The adverse effect of diabetes known as diabetic retinopathy could result in blindness when identification occurred past the early stage. An automated detection system proved essential because manual screening required extensive time and resources at health- care facilities. The proposed research established a deep learning-based solution for quick diabetic retinopathy screening, where the system identified and categorized diabetic retinopathy symptoms into five different stages.Born of the necessity of efficient screening, the model made use of hybrid transformer-based architecture, which combines EfficientNetB0 to extract initial features along with Swin Transformer to learn hierarchical self-attention. This fusion can be used to improve the contextual interpreta- tion of fundus images with several severity grades of diabetic retinopathy. To even increase the prediction performance, attention based ensemble approach was employed producing stronger classification.The APTOS 2019 Blindness Detection Dataset was used in training and testing of the model since it has five graded levels of the disease. Data showed great class imbalance; therefore, classical resampling methods were used in the first place. To add to the imbalance and to enable generalization, it was even at- tempted to make use of GAN-based data augmentation strategy.The most effective model, EfficientNetB0 + Swin Transformer using attention-based ensemble and GAN augmentation, showed the highest accuracy of 0.85, a micro F1-score of 0.85 and a macro F1-score of 0.63, allowing one to differentiate between all of the five stages of diabetic retinopathy quite successfully. The addition of Grad-CAM visualizations further improved system interpretability, e.g., because it illustrated regions on retinal images that were relevant to the disease, thereby providing clinicians better insight into system decisions. The offered procedure was also effective to provide a reliable mechanism to identify the early-stage diabetic retinopathy in the most effective and non-opaque manner, which is beneficial to patient services by reducing the risk of blindness.

Zitieren

APA

Sahithi, B. K., Chandrika, A. S., & Venugopalan, M. (2026). EfficientNet and Swin Transformer Fusion for Diabetic Retinopathy Screening. Procedia Computer Science, 283, 4560-4569. https://doi.org/10.1016/j.procs.2026.06.512

Vancouver

Sahithi BK, Chandrika AS, Venugopalan M. EfficientNet and Swin Transformer Fusion for Diabetic Retinopathy Screening. Procedia Computer Science. 2026;283:4560-4569. doi:10.1016/j.procs.2026.06.512

BibTeX
@article{Sahithi2026,
  author = {Sahithi, B. krishna and Chandrika, Achalla Sruthi and Venugopalan, Manju},
  title = {EfficientNet and Swin Transformer Fusion for Diabetic Retinopathy Screening},
  journal = {Procedia Computer Science},
  year = {2026},
  volume = {283},
  pages = {4560--4569},
  doi = {10.1016/j.procs.2026.06.512}
}
Kostenlos starten →