Ensembling Convolutional and Transformer Architectures for Retinopathy of Prematurity Screening in Neonatal Fundus Image

Overview

Neonatal eye care faces significant challenges due to diseases like Retinopathy of Prematurity(ROP). Accurate and punctual detection of ROP is crucial for the preservation of infant eyesight. This paper proposes an approach to infantile retinal disease detection by leveraging advanced Convolutional Neural Networks(CNN) & Vision Transformers(ViT). Specifically, we utilize MobileNet-V2, a lightweight deep convolutional network architecture that prioritizes low computational cost, and ResNet-50, a widely adopted deep learning network that enables deeper architectures. In addition, we employ SWIN-Transformers, a hierarchical vision transformer that uses self-attention-based shifted windows. Comparative analysis was performed on fundus images across 3 classes of retinal diseases. The results show that the SWIN-Transformer performs well on its own with an accuracy of 77%, outperforming MobileNet-V2 and ResNet-50. In addition, a hybrid ensemble SWIN-Transformer and ResNet-50 model achieves an accuracy of 84% with a high degree of precision (81%) and recall (87%).The research findings of this study highlight the implications of deep learning ensemble CNN and ViT models in screening performance, while maintaining interpretability through class activation analysis.

Competition history

• CSEF 2026 Computational Science (Senior Division) · Entry S-07-02

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Computational Sciences projects
• CSEF Computational Sciences projects
• Every CSEF 2026 project