Objective Vision loss is caused by age-related macular degeneration because of soft drusen, diabetic macular edema and choroidal neovascular disease. Early detection and treatment of these fundus diseases have emerged as a major health concern for all countries. Professional doctors often use retinal images from optical coherence tomography (OCT) to diagnose eye diseases. However, because there are several types of retinopathy images and the lesion area is similar, manually classifying OCT retina images is a time-consuming and difficult task. With the development of artificial intelligence, researchers began classifying medical images using classic machine learning algorithms and deep learning in its branch areas, eventually progressing to the automatic classification of OCT retinal images. Several researchers are only concerned with classification accuracy and ignore the possibility of clinical application. Consequently, the network model’s parameter amount, computational complexity and floating-point operations (FLOPs) calculation amount are increasing, and the model is making inference predictions, which consumes a long time to complete. In this paper, a multi-channel, multi-scale lightweight convolutional neural network is proposed to automatically classify OCT retinal images for achieving high ophthalmic disease classification accuracy. In the future, doctors will be able to quickly view detection results in the clinic.
Methods In this study, a multi-channel OCT retinal image automatic classification deep neural network is used. The neural network model is based on the GM-OCTnet algorithm, which includes a light quantum spatial attention mechanism distinct from the convolution operator and a lightweight convolution block to replace the two modules in the original model for automatically classifying OCT retinal images. Image pre-processing, dataset division and classification using the model algorithm are the steps taken to achieve the automatic classification of the entire OCT retina. First, image cropping is performed on the collected OCT image, the blank area of the OCT image is cropped, the marginal blank area is filled and bilateral filter denoising and other pre-processing methods are used to overcome the interference of image background noise on the classification accuracy. Then, the pre-processed image is divided into three sets: training, validation, and test sets. Afterwards, the OCT images from the training set are trained using the proposed multi-channel lightweight deep neural network GM-OCTnet model algorithm. Following the test, the well-trained model is used to classify unclassified retinal images automatically. In addition, the results of this work on the automatic classification of OCT retinal images are validated by comparing the proposed model with three traditional lightweight models on the OCT data set, and different data sets are used to further validate the algorithm’s performance.
Results and Discussions Different pre-processing methods were used to process the OCT images after evaluating the quality of two different data sets (Figs. 4 and 5). The experimental results show that when the number of groups is 4, the proposed multi-channel OCT automatic retina classification network achieves an average accuracy of 96.1%, which is 2.6% higher than that of the original model GhostNet in the automatic classification of OCT retina images. Its file size is 2.64×10 6smaller than that of MobileNetV3. The training and verification accuracies of the GM-OCTnet model when the grouping g=4 increase gradually with the increase of the training period and tend to stabilise, based on the relationship between the verification loss rate and the verification accuracy rate and the training loss rate and the training accuracy rate curve. When comparing the 50 training processes of different models, the proposed model is found to exhibit a higher accuracy rate than other models when the number of groups is equal to 4 and prioritises reaching the best value (Fig. 6). Overall, the model algorithm proposed in this paper has achieved high accuracy in the automatic classification of OCT retinal images. In addition, when the experimental results of two different datasets are compared, it is discovered that the automatic classification of datasets 1 and 2 achieved high classification accuracy using this algorithm (Tables 1 and 2).
Conclusions This study proposes a multi-channel, multi-scale lightweight network for automatically classifying OCT retinal images. The effect of the lightweight neural network GM-OCTnet based on the OCT image datasets in classifying and diagnosing the four types of ophthalmic conditions, i.e. choroidal neovascular disease(CNV), diabetic macular edema(DME), drusen and normal patients, were tested and evaluated. Two different datasets are used to further validate the performance of the algorithm proposed in this work. To validate the effectiveness of the GM-OCTnet model for OCT image classification, accuracy, parameter amount, calculation amount and weight file size are used as evaluation criteria. It is found the proposed OCT classification model has improved classification accuracy through experimental results. When used in the clinic, it can improve professional ophthalmologists’ diagnosis efficiency for patients with ophthalmic diseases and also reduce missed diagnosis and misdiagnosis of patients.
