Research On Retinal OCT Image Classification Method Based On Deep Learning

Retinal diseases have greatly affected people’s daily lives.With the increasing number of patients with retinal diseases,ophthalmologists are gradually unable to meet the needs of retinal disease patients through manual diagnosis.At present,deep learning methods have been widely applied in the field of medical assisted diagnosis.By using deep learning methods,ophthalmologists can diagnose retinal disease patients more efficiently and accurately.However,when using deep learning to classify retinal OCT images,there are problems such as weak network feature extraction ability and inability to achieve ideal classification accuracy when the number of training samples is insufficient.In response to the above issues,this article conducts research on retinal OCT image classification methods based on deep learning.The specific research content is as follows:In response to the problem of unsatisfactory classification performance when processing the original retinal OCT image dataset,this article preprocesses the original retinal OCT image data,including data equalization,image denoising,image normalization and standardization,to accelerate the training speed of the network,improve the generalization ability of the network,and make the network training more stable.In response to the problem of weak feature extraction ability in some existing networks,this paper proposes a multi-scale residual network.The network uses ResNet50 as the baseline network,combined with convolutional block based attention mechanism,to enhance the network’s ability to extract important information of input image distribution in channels and spaces.At the same time,a multi-scale residual module is used to achieve identity mapping while the network can achieve identity mapping,Effective feature extraction of retinal disease information at different scales and effective learning of small disease features in images.In view of the problem that deep learning cannot achieve the ideal classification accuracy when the number of training samples is insufficient,an improvement is made on the basis of the above network,and a recursive residual module is designed.Through traversing the input information of the network,the network can obtain better classification results when using the same input.At the same time,the Dropout method is used in the network to prevent overfitting phenomenon in the training process.Finally,The multi-scale recursive residual network is fused with the transfer learning method.First,pre training is carried out on the data set with sufficient data,and then the pre trained model is migrated to the small sample data set to effectively solve the problem of insufficient training samples.On the dataset OCT2017 with sufficient sample size,The multi-scale residual network used in this article has a final accuracy of 97.3%,which is 2.2% higher than the baseline network.When facing the SN dataset with insufficient samples,the accuracy of the multi-scale recursive residual network proposed in this paper reaches 82.5%,which is 2.9% higher than that of the multi-scale residual network.Finally,after the transfer learning method is cited,the classification accuracy on the small sample dataset reaches90%,fully verifying the effectiveness of the method used in this paper.