Research On Classification Of Macular Diseases And Measurement Of Macular Hole Based On Deep Network

The macular is located at the centre of the fundus retina and serves as a provider of visual function.When the macular is diseased,it will damage the vision and even cause blindness.Therefore,early screening and timely intervention for macular diseases are essential.Different macular diseases show other characteristics in optical coherence tomography(OCT)images.Ophthalmologists use professional knowledge and experience to determine the type and severity of macular diseases,and take appropriate treatment measures.However,this purely artificial diagnosis of macular disease is timeconsuming,labor-intensive,and inefficient.The development of deep learning technology has made it possible to diagnose eye diseases with computer assistance.In this paper,the research on macular disease classification and hole size measurement based on a deep neural network algorithm is carried out on retinal macular OCT images.With the aid of the deep learning method,the features of macular OCT images are automatically learned to distinguish the type and severity of macular diseases,assist ophthalmologists in diagnosing and treating macular diseases,improve their work efficiency,and achieve an intelligent diagnosis of eye diseases.The contributions of this article mainly include the following two aspects.1)Aiming at the problem that the small lesion area and unclear lesion characteristics of macular diseases make it difficult to identify the types of macular diseases,this paper proposes an improved convolutional neural network model for the automatic classification of macular diseases.The improvement idea of this model is divided into three points.The first point is to add a multi-scale feature fusion module based on the pure convolutional neural network model to splice and fuse feature images with different receptive fields,increasing the nonlinear expression ability of the model,and thereby enabling more abundant pathological features to be extracted.The second point is to reduce the attention paid to redundant features of macular OCT images,enhance the attention paid to features of lesion areas,and reduce the impact of redundant features on classification performance by increasing attention mechanisms.The third point is to train the model through the weighted loss function of effective samples,thoroughly learn the pathological features of small sample categories,solve the problem of data sample imbalance,and improve the classification performance of small sample macular disease categories.Relevant experimental results show that the improved neural network model improves the overall accuracy of macular disease classification,improves diagnosis and treatment efficiency,and reduces misdiagnosis and missed diagnosis.2)To solve the problem of large noise and high complexity in macular hole OCT images,an algorithm with three stages for measuring macular hole size is proposed in this paper.Firstly,the improved U-Net model is used to achieve accurate segmentation of the macular region,while the residual network model is used to achieve the classification of macular categories,divided explicitly into the normal macular and the macular hole.Secondly,edge detection and noise removal are performed on the segmented image of the macular hole to obtain the critical edges of the macular hole.Finally,the left and right endpoints of the macular hole are determined based on the gradient variation in the critical edges of the macular hole,thereby measuring the size of the macular hole.The relevant experimental results prove that the error between the measurement results of the proposed three-stage algorithm and the manual measurement results of professional doctors is within an acceptable error,and can quickly provide a more accurate and unique indicator of hole size for the OCT system.