In this paper, a novel architecture of Deep Belief Networks is proposed for the automatic classification of two types of macular edema. The method is based on two stages: in the first stage, the Deep Belief Network is trained; in the second stage, using pre-trained weights a multilayer perceptron network is initialized and retrained for further tuning. The database consists of annotated fundus images including two types of edema: Drusen and Cotton Wool Spots. The database comprises 140 healthy samples, 140 samples with Drusen, and 70 samples with Cotton Wool Spots, all manually labeled by a specialist from the Mexican Social Security Institute. To balance the number of samples per class, ADASYN (Adaptive Synthetic Sampling Approach for Imbalanced Learning) was used along with data augmentation strategies. The model was evaluated in terms of accuracy, precision, recall, F1-score, and time per image for MobileNetV2, ResNet50, DenseNet, and EfficientNetB7 models. The results showed that the proposed architecture outperformed other state-of-the-art methods, achieving and metrics of 0.8286, 0.8273, 0.8286 and 0.8265 respectively. Furthermore, the computation time per image was 0.0026 s—approximately 379 times faster than the second-best performing model. This highlights the effectiveness of the proposed method by providing a highly efficient experimentation framework that not only accelerates the development process but also ensures easy adaptability for other applications.