Research On Cataract Screening Method Based On Artificial Intelligence

Lens slit lamp image is an important basis for the diagnosis of cataract.Experienced doctors often observe the lens slit lamp image according to their own experience to determine the severity of the cataract.The traditional image-based cataract classification method first needs to manually design features and then perform classifier training,which is less efficient and subjective.To this end,this thesis designs an automatic learning model of cataract features based on convolutional neural network,which can classify lens slit lamp images according to the severity of cataract,and is generally applicable to nuclear cataract,cortical cataract and posterior subcapsular cataract,which can provide better auxiliary information for the diagnosis of clinical cataract.The research content of this thesis is mainly divided into four parts.In order to solve the current problem of lack of cataract dataset,this subject has collected a large number of images of lens slit lamp with different types of cataracts,the ophthalmologist divides it into normal lens,early cataract lens and cataract lens according to their severity to construct the MSLPP dataset;The image in the MSLPP dataset is preprocessed,and the OTSU-Hough method is used to locate and segment the pupil part,and then the segmented image is subjected to illumination enhancement and data augmentation processing;In order to solve the problem of automatically extracting depth features and classification,this thesis uses ImageNet pre-trained Inception-V3 model and parameters as the initial model,and adopts the idea of transfer learning.After changing the fully connected layer,continue training with MSLPP dataset,using the method of cross-validation to continuously test the accuracy of the model to obtain the final classification model;Finally,this thesis brings together all the previous results,developed a mobile phone client-based cataract screening system.The doctor can use the mobile phone equipped with the application system to collect and upload the lens slit lamp image with the mobile phone slit lamp.The system will preprocess the uploaded image and automatically feedback the cataract,early cataract and normal conditions,which has strong practicability.In this thesis,the characteristics collected by the cataract screening model were visual analysis,and the reliability of the model was evaluated from the four perspectives of Accuracy,Recall,Precision and F1-meature.In order to prove that the cataract screening model is not affected by light,this thesis compares the model in bright and dark environments.The experimental results show that the accuracy of model recognition is 94.84%,which effectively realizes the classification of cataract images and can provide an auxiliary diagnosis basis for ophthalmologists.