Fundus Image Screening for Diabetic Retinopathy

Abstract

Objective Diabetes is a worldwide chronic disease, which can cause changes in vascular performance and result in complications such as diabetic retinopathy (DR). If a patient has DR and it goes undetected, the patient's eyesight may be lost. Therefore, early detection and early treatment of DR is important for reducing the blindness rate in various countries. At present, the acquisition of fundus images mainly depends on nonmydriasis color fundus photography, which can clearly capture the soft and hard exudates, bleeding points, microvessels, and so on. Currently, the ophthalmologists who observe fundus images have to endure heavy workloads and massive job-related stress owing to problems such as insufficient expert resources. To reduce their workload, it is necessary to use computer-aided diagnosis. Moreover, computer-aided diagnosis is more accurate. In the field of medical imaging, using classical machine learning and deep learning to classify medical images has become key research areas and gradually the subfield of retinal fundus image classification has developed. However, there is a serious problem with machine learning and deep learning algorithms for fundus image classification developed by many researchers, that is, they continuously increase the complexity of the model to pursue high accuracy. This results in a corresponding increase in computational complexity in terms of the number of floating-point operations and the size of the parameters of the network model, thus reducing the speed and increasing the memory utilization. An inefficient classification model is less likely to be used in clinical practice. The purpose of this paper is to propose a simple network model. Compared with the most advanced model, our model has not only high accuracy, precision, and sensitivity but also high speed. More importantly, it has the potential to be used in clinical practice. Methods We improved the network architecture of the RepVGG model proposed by the Kuangshi' s group and proposed a novel model Channel Attention-RepVGG (CA-RepVGG). The RepVGG with simple structure was used to replace the complex module as the main part of the classification model, and efficient channel attention was selected to replace the squeeze-and-excitation (SE) for a good classification of images related to DR. Then, CA-RepVGG was tested on the new dataset of DR images. The main research includes the following. First, the multibranch architecture employed during training and the single-branch architecture employed during inference were decoupled using the structural reparameterization method, which greatly reduced the complexity of the model and met the requirements of simplicity in structure. Second, a new lightweight attention module was used to improve the performance of the convolutional neural network and enhance the ability of feature extraction from retinal fundus images. Finally, the parameters, speed, precision, accuracy, and sensitivity of several classical networks in image classification were compared. Results and Discussions The proposed model is tested on 1096 pictures of dataset 1 and 500 pictures of dataset 2. The accuracy of dataset 1 is 92. 4%, the precision is 91. 6%, and the sensitivity is 96. 5%. For dataset 2, the accuracy is 93.9%, the precision is 96. 3%, and the sensitivity is 93. 8%. The confusion matrix of dataset 1 is shown in Fig. 6 and that of dataset 2 is shown in Fig. 7. The two figures show the classification results of the pictures; the values on the diagonal present the number of patients correctly classified by the model. Only a few images have errors, the overdiagnosis distributes on the lower left of the diagonal in the image and the missed diagnosis distributes on the upper right of the diagonal. The confusion matrices show the superiority of the proposed model. Furthermore, CA-RepVGG has the fastest speed, although it has the largest parameters in the compared models. Owing to the single-branch architecture and 3x3 convolution, our model is not complex. The CA-RepVGG can process 415 pictures per second, which is 15.3% higher than those by ResNet-50. Conclusions CA-RepVGG can be used in clinical practice. The simplicity of the model and the small amount of calculation ensure the feasibility and reliability of CA-RepVGG. In this paper, CA-RepVGG is used to test and evaluate the classification effect of DR images in two datasets. At the same time, VGG-16, Inception-V3, ResNet-50, and ResNext-50 are compared with our model, and the accuracy, precision, and sensitivity of the network demonstrate the advanced nature of our model. The experimental results show that the model is not only feasible but also superior in classification. In the future, if our proposed model is applied to clinical practice, it can enhance the diagnostic efficiency of professional ophthalmologists regarding ophthalmic diseases, especially in remote and poor areas, ensuring that more patients can be treated in time and avoid losing their eyesight. If more datasets can be used to train the model in the future, the accuracy of automatic classification can be further enhanced and better results can be achieved in clinical practice.