The study proposes a CheXNetCBAM model for diabetic retinopathy classification from fundus images, based on DenseNet121 with an added CBAM module for better feature representation through channel and spatial attention mechanisms. Grad-CAM is also used to visualize areas important to model predictions. The model is compared with CheXNet, DenseNet121, MobileNetV2, VGG19 and ResNet50 architectures on APTOS 2019 and DDR datasets. It achieves an accuracy of 96.12% on APTOS 2019 and 96.33% on DDR, thus surpassing the compared models. Incorporating CBAM improves discriminative feature learning. Further prospective evaluation and external validation are needed for clinical use.