Diabetic retinopathy (DR) is a major cause of vision loss in diabetic individuals, and early identification is essential to prevent progression to blindness. This work presents a machine learning approach using convolutional neural networks (CNN) to detect DR severity automatically. The model is designed to extract detailed features from retinal images by using a multi-scale feature fusion network built on MobileNetV3 using machine learning. Experimental results indicate that this approach surpasses existing methods, including various machine learning parameters, achieving high accuracy, sensitivity, and specificity. The proposed method starts by taking a retinal image as input. It then extracts three key features that help identify diabetic retinopathy. After that, the model calculates each feature’s importance (weight) and merges them using a weighted fusion process to improve accuracy. Finally, the classifier analyzes the refined data and determines the severity of DR. The experimental results highlight its potential to revolutionize DR screening by offering faster, more precise, and widely accessible diagnosis, particularly for remote and underserved regions. The approach facilitates prompt and accurate identification of diabetic retinopathy, supporting early intervention and enhanced patient outcomes.