<p>In color fundus photography (CFP), pathological features often manifest as a combination of localized, subtle lesions and global structural alterations. Consequently, it is crucial to jointly capture fine-grained local details and comprehensive global information for computer-aided diagnosis of ophthalmic diseases. In the field of computer vision, convolutional neural networks (CNNs) excel at capturing local patterns and spatial hierarchies, whereas Vision Transformers (ViTs) demonstrate superior capability in modeling long-range dependencies and global contextual relationships. In recent years, ophthalmic foundation models have shown remarkable transferability and generalization across diverse downstream tasks. Inspired by these advances, we propose the PDNet, a parallel dual-branch framework that integrates a ViT with a CNN model to enable collaborative modeling of local and global representations. Moreover, to alleviate the loss of critical shallow information in deeper layers of the CNN branch, we introduce the progressive cross-stage feature enhancement (PCFE) module, which progressively propagates shallow features across network stages to incrementally enhance feature representations of deeper layers. Experiments conducted on multiple public datasets demonstrate the effectiveness of PDNet, achieving good performance on ophthalmic disease classification tasks, including both glaucoma diagnosis and diabetic retinopathy grading.</p>