RetiNet-X: A cross-branch deep fusion architecture for diabetic retinopathy detection

Diabetic retinopathy (DR) is a leading cause of preventable vision loss, underscoring the need for reliable automated screening systems. While deep convolutional networks have achieved promising results, many existing approaches rely on single-stream feature learning, which limits representation diversity and reduces sensitivity to subtle lesion patterns. In addition, computationally intensive backbones do not always guarantee improved generalization across datasets. This paper presents RetiNet-X, a novel dual-branch deep learning framework for binary DR classification. The proposed architecture enhances hierarchical feature representation by integrating a Kernel-Refined and Depth-Enhanced Inception module with spatial attention to impose lesion-aware inductive bias. A lightweight transfer learning branch extracts complementary high-level semantic embeddings. These heterogeneous features are fused to improve discriminative separability and robustness. Experimental evaluation on the APTOS-2019 and Kaggle_DR benchmark datasets demonstrates that RetiNet-X achieves accuracies of 97.81% and 96.5%, respectively, outperforming contemporary methods. The proposed framework offers an efficient and scalable solution for automated DR screening in clinical and tele-ophthalmology environments.