Abstract: To address the challenges of uneven inter-class distribution and difficulty in lesion area recognition in retinal fundus image datasets, this paper proposes a fusion dual-attention retinal disease grading algorithm with PVTv2 and DenseNet121. First, retinal images are preliminarily processed through a dual-branch network of PVTv2 and DenseNet121 to extract global and local information. Next, spatial-channel synergistic attention modules and multi-frequency multi-scale attention modules are applied to PVTv2 and DenseNet121, respectively. These modules refine local feature details, highlight subtle lesion features, and enhance the model's sensitivity to complex micro-lesions and its spatial perception of lesions areas. Subsequently, a neuron cross-fusion module is designed to establish long-range dependencies between the macroscopic layout and microscopic texture information of lesion areas, thereby improving the accuracy of retinal disease grading. Finally, a hybrid loss function is employed to mitigate the imbalance in model attention across grades caused by uneven sample distribution. Experimental validation on the IDRID and APTOS 2019 datasets yields quadratic weighted kappa scores of 90.68% and 90.35%, respectively. The accuracy on the IDRID dataset and the area under the ROC curve on the APTOS 2019 dataset reached 80.58% and 93.22%, respectively. The experimental results demonstrate that the proposed algorithm holds significant potential for application in retinal disease grading.