Hybrid Graph-Convolutional Architectures for Diabetic Retinopathy Screening: The SCG-ARMA-EfficientNet Framework

Abstract

Diabetic retinopathy (DR) grading from fundus images remains challenging due to noise artifacts, class imbalance, and subtle inter-class differences. This paper presents an ensemble deep learning framework that integrates a Self-Constructing Graph (SCG) module with adaptive image preprocessing for robust DR classification. The proposed SCG layer, embedded within an EfficientNetV2L backbone, dynamically infers relational structures between extracted features through stochastic covariance estimation, enabling the model to learn latent anatomical dependencies in retinal images. Complementing this, a Master-Slave Adaptive Notch Filter (MSANF) automatically adjusts its denoising parameters based on local noise levels, enhancing low-quality images without manual intervention. To mitigate class imbalance, we introduce a targeted augmentation strategy that synthesizes minority-class samples using conditioned geometric transformations. The system combines SCG-enhanced EfficientNetV2L with ResNet50 and DenseNet121 into an ensemble, achieving an 84% balanced accuracy on the APTOS 2019 dataset—outperforming baseline models by 4.2%. Ablation studies validate the SCG’s role in modeling feature relationships (↑2.8% F1-score) and MSANF’s impact on noisy inputs (↑3.5% sensitivity). This work advances explainable DR diagnosis by jointly addressing data quality and hierarchical feature learning, with potential for deployment in screening pipelines.