<![CDATA[Diabetic Retinopathy (DR) is a cause of blindness. Early detection has the potential to save the patient's vision. Reading fundoscopic photos requires more expertise and effort by the ophthalmologist. There are many visual similarities in lesions and only minor differences in the spatial domain. A computer-assisted automatic detection system is needed to assist medical experts in DR diagnosis and can reduce costs. This study proposes a combination method of Convolutional Neural Network (CNN) and Support Vector Machine (SVM) for the automatic classification of Diabetic Retinopathy (DR). The pre-train architecture Inception-V3 uses for feature extraction of input data. After training and getting the best model, the next is classification with SVM. Data augmentation techniques use to multiply and add variations to the dataset. Before the feature extraction stage, the dataset will process by separating the green channel from the RGB image. Next, the CLAHE will require increasing the contrast of the picture. This study aims to improve the performance in multi-class DR classification. The proposed model uses four classes of unbalanced and small datasets from retinal fundus images. This paper also compares the combined performance of CNN SVM with CNN Softmax based on the accuracy value to validate the results. Our experiments show that the combination of CNN SVM can increase the accuracy of auto-detection of DR severity up to 11.48% better than CNN softmax. The results showed that the pre-trained architectural model from the combination of Inception-V3 with SVM classification improves the accuracy extensively, even on small and unbalanced datasets.]]>
