Diabetic Macular Edema (DME) refers to hard exudate (HE) lesions close to the macular area of the retina. Furthermore, it is a condition that leads to vision impairment and can potentially cause total blindness. Detection and timely intervention can potentially provide a curative effect for this condition. The early-stage identification of DME presents challenges and is susceptible to errors. This paper introduces a new approach to identifying the disease’s presence and is classified into three stages (normal, mild, and severe). This paper comprises five stages: Stage one includes preprocessing by improving the quality of the images using improved CLAHE histogram equalization, which improves the contrast of images. In the second stage, I used the proposed CNN to classify whether the eyes were healthy or infected. If infected, the disease’s severity is decided in the third stage by detecting the optic disk (OD) and then segmentation because it contains a color density similar to the color density of the disease in hard exudate. In the fourth stage, determine the degree of seriousness of the injury by extracting the frequency features resulting from the Wavelet Scattering Transform (WST). In the final stage, we combined deep features from (WST) with the SVM classification process, mild or severe, using MESSIDOR and IDRiD datasets. In the first stage, the CNN model attained 99.7% and 98.9%, respectively; in the second stage, 99.7% accuracy was attained with the YOLO model in MESSIDOR and generalized to the rest of the datasets. In the third stage, the combined deep features with the SVM classifier achieved an accuracy of 99.5% and 99.1%, respectively. The proposed method’s deep features significantly impact finding differences between data, which greatly helped classification and increased the system’s accuracy.