İsa Ataş
Automated grading of Diabetic Retinopathy (DR) remains a challenging computer vision task due to the need to jointly model fine-grained local lesions and global structural patterns in retinal images. Existing hybrid architectures combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) typically rely on static feature fusion strategies, which limit adaptability to varying image characteristics and often result in feature misalignment. To overcome this limitation, this study proposes Reti-TransNet, a dynamic hybrid framework that integrates an EfficientNet-B0 backbone with a Swin Transformer branch through an Adaptive Gated Fusion (AGF) mechanism. The AGF module formulates feature fusion as a control problem and employs a learnable gating network to dynamically recalibrate channel-wise contributions of local and global representations. In addition, a Nelder–Mead optimization scheme is applied to fine-tune decision thresholds, directly maximizing the Quadratic Weighted Kappa (QWK) metric to preserve ordinal consistency. Experimental results on the APTOS 2019 dataset show that the proposed approach achieves a QWK score of 0.905 (95% CI: 0.88–0.93) and a multiclass accuracy of 84.31% in the primary configuration (mean QWK of 0.897 ± 0.006 and mean accuracy of 79.94% ± 3.21% across five independent runs), outperforming representative static fusion baselines. Further evaluation on the external IDRiD dataset indicates favorable primary screening performance under domain shift, achieving an AUC of 0.963 for healthy class detection, although detailed multiclass generalization remains limited due to sensitivity drops in early-stage disease detection. The proposed dynamic fusion strategy offers a promising and resource-efficient approach for computer-aided DR diagnosis.