<p>Ocular OCT images of diabetic retinas serve as a foundation for ophthalmologists to assess the presence and severity of lesions. Leveraging deep learning methods for automated diagnosis and analysis of diabetic retinopathy can reduce workload, minimize human error, and decrease misdiagnosis rates. This paper proposed an SSD-based target detection network to predict whether the retina had lesions or not, by uniformly preprocessing the ocular OCT images collected from the Second Hospital of Shanxi Medical University, labeling the lesion areas to construct an ocular OCT diabetic retinopathy dataset, and inputting it into the SSD target detection network to extract image features, thereby realizing the detection of diabetic retinopathy. The transfer learning method and stochastic gradient descent algorithm were applied to optimize the SSD network model to enhance lesion classification accuracy and detection. The proposed method was compared with the YOLOV5 and YOLOV7 models. The experimental results showed that the SSD target detection network had high detection accuracy in the hierarchical detection of diabetic retinopathy. The average accuracy of the model on the verification set was 95.72%, and the F1 score was 0.947. This study has important implications for the screening and treatment of patients with early sugar reticulum.</p>