Automatic Cataract Classification Based On Deep Learning

Cataracts are caused by the accumulation of proteins in the lens,one of the most serious eye diseases leading to blindness.Early detection and treatment can reduce the suffering of patients and prevent visual impairment from becoming blind.In the medical field,the use of medical image processing and analysis as a means of auxiliary diagnosis and treatment to detect cataract as soon as possible that has important research value.This article reports the current domestic and international research status concerning the automatic cataracts detection and classification.The current research on cataract auto-diagnosis usually based on a set of pre-defined features and then classifying them using classification algorithms,but this kind of pre-defined feature sets may be atypical,incomplete,or redundant.This paper studies the application of deep learning in cataract auto-classification task,and proposes a new cataract auto-classification method based on a hybrid feature extraction model.The work accomplished in this paper is as follows:(1)Using the deep convolutional neural network(CNN)to automatically extract the feature set of the complete fundus image and construct a cataract auto-classification model based on global features.Experiments show that the feature set extracted automatically by using the Alexnet network model can provide better cataract feature representation than the existing pre-defined feature set.(2)A method is defined to perform interpretability analysis of the global features of the fundus images automatically extracted by the CNN classification model.Using deconvolution neural network(DN)method to visualize the characteristics of each layer of the CNN global network model,revealing the response relationship between the highest active feature map of any layer in the model and the input,and observing the characteristics from low level to high level abstraction Hierarchical transformation.It has been experimentally verified that relying solely on the global feature set of fundus images is not the best choice for identifying the severity of cataract.(3)Building a cataract classification model based on local features.In order to overcome the drawbacks of the CNN classification model based on global features,we propose a local CNN classification model by constructing a variant data set to extract the local detail features in the fundus image,highlighting the vein structure of the blood vessel and The details of the tertiary blood vessels.(4)Based on ensemble learning,a global cataract auto-classifier based on global-local hybrid feature extraction model was designed by integrating global and local feature-based network models.The experimental data used in this paper is a real set of clinical fundus images,and verified by the cross-experiment results,the method proposed in this paper has achieved 86.24% accuracy in cataract auto-classification task,which is better than the existing methods.