VGG-ST: A VGG-Swin Transformer-Based Model for ROP Disease Diagnosis

The detection of Referral-Warranted Retinopathy of Prematurity (RW-ROP) is crucial for preventing severe visual impairment in premature infants. Recent studies have demonstrated that deep learning models, particularly CNNs, are effective in classifying ROP. However, the comprehensive extraction and integration of ROP-relevant features from retinal images for accurate classification remains challenging. In this paper, we propose a hybrid model based on Very Deep Convolutional Networks (VGG) and Swin Transformer (VGG-ST) for identifying RW-ROP using retinal images. The VGG-ST model first employs the VGG19 architecture to extract detailed local image features and the Swin Transformer V2 architecture to capture comprehensive global contextual information. It then introduces a novel feature enhancement module that combines these local and global features through an adaptive integration strategy, optimizing the feature representation for more accurate RW-ROP detection. Finally, the integrated features are processed through a classification module to predict the probability of RW-ROP, distinguishing between normal and RW-ROP cases. We also present customized data preprocessing techniques to address class imbalance and retinal image blur issues inherent in the e-ROP dataset. Experimental results demonstrate that the VGG-ST model offers improved classification sensitivity compared to existing methods, making it a promising tool for automated ROP screening in clinical settings. The source code is available at https://github.com/hawk-sudo/VGG-ST.