Application Research Of Convolutional Neural Network In Ocular Medical Image Classification

With the improvement of medical imaging equipment in the past decade,a large amount of medical image data has been accumulated,which has promoted the development of medical image processing,and the eye as an important medium for human beings to receive and perceive external visual information,the processing and analysis of its corresponding eye medical images is also an important research direction.With the continuous improvement of deep learning theory and the continuous updating of hardware facilities,deep learning technology has been successfully applied in the fields of image classification,object detection,semantic segmentation,etc.Among them,the convolutional neural network algorithm is the most prominent among many algorithms.Therefore,applying deep learning technology to automated analysis of ocular medical images can not only reduce the doctor’s heavy medical image screening work,improve work efficiency,but also help patients prevent disease and timely treatment,which is of great practical significance.In this paper,deep learning is applied to two types of ocular medical images,one of which is the image of thyroid-associated eye disease-the eyelid image,and the other is the image of diabetic retinopathy-the fundus image.The main contents of this paper are as follows:(1)This paper briefly introduces the development process of deep learning,common algorithms and training methods,and focuses on the structure,principle and common network model of convolutional neural networks.Finally,the commonly used deep learning framework and the application area of deep learning are expounded.(2)Classification of thyroid associate ophthalmopathy(TAO).Firstly,the disease region of TAO image is obtained by data preprocessing operation.Then,based on the characteristics of multi-bit map,a composite convolutional neural network model is designed.The model contains three sub-networks,and extracts the features of corresponding bitmaps and classifies them.The structure of the network is based on the improved LeNet-5 network,and multiple sets of experiments on the network structure and parameters are explored,and then compared with the original LeNet-5 network.Finally,the results of the final inspection process are obtained by integrating the output of the three sub-networks.(3)Classification of the extent of diabetic retinopathy(DR).In terms of data sets,preprocessing operations are used to eliminate image noise,and cost-sensitive learning and resampling methods are used to solve sample imbalance problems.On the network side,a deep supervision Inception-Residual network-DSIRNet was designed,and a deep monitoring method was introduced to assist the training.Finally,the effectiveness of the proposed network and method is verified by comparative experiments.In summary,this paper mainly studies the feature extraction and classification of eye medical images based on convolution neural network model.Experiments verify the validity and practical application value of the proposed model and method.It also has a good reference value for other types of medical image classification.