High risk of diabetic retinopathy which, if not diagnosed on time, may lead to complete blindness, exists in the long-term diabetic patients. This paper presents a method using multimodal fusion technique utilizing important information from retinal fundus images by combing features from two prominent pre-trained models, VGG19 and ResNet50 as well as other popular architectures being NASNetLarge, EfficientNet, DenseNet121, MobileNetV2, InceptionV2, and Xception. This model utilizes cross-pooling fusion to capture fine image details that are required for the precise diagnosis of DR. Of the various combinations of pre-trained models, the combination of VGG19 and ResNet50 features proved the most successful in improving the model’s ability to detect minute changes in retinal pictures and categorize the severity of DR levels accurately. Neural network architecture has been used for classification. It found DR severity levels at a high accuracy rate of 83.33% along the normal, mild, and moderate stages. Other key metrics in assessing performance include ROC curve, accuracy, and confusion matrix. Very impressive results have been recorded as this method was experimented with on the Kaggle APTOS 2019 benchmark dataset. The encouraging accuracy of this fusion-based model underscores its potential for early DR intervention, which is quite important for diabetic patients for preserving their vision.