Automatic Glaucoma Classification Research Based On Transfer Learning

Ophthalmic disease occurs in the visual system of the human body,including the eyeball and related parts.This disease is divided into many categories,and glaucoma has a serious impact as a representative.The development of glaucoma is often continued.If it can be detected early,it can effectively prevent from deteriorating the disease condition,which could lead to visual impairment and even blindness.Due to the complexity of glaucoma itself,early diagnosis always requires a large amount of prior knowledge of ophthalmologists as the basis,which will cause certain pressure on both doctors and patients.Many scholars and research institutions are now actively taking researches on the prediction model of glaucoma.However,there are two problems as follows: on the one hand,since most of the researches use the image processing method to define features artificially in advance,the feature selection may not conform to the expression of real data.In other words,this feature set may not be truly objective and sufficient.On the other hand,due to the limitations of its own circulation channels,it is difficult to obtain a large number of complete annotated data for medical data as model training.It is a serious constraint for model construction,especially in deep learning.Thus,how to use a small amount of medical data cost to train a reliable prediction model is also a problem to be solved urgently.In order to figure out these problems,we proposed a new glaucoma recognition algorithm based on deep transfer learning,which aims at knowledge sharing between glaucoma and cataract digital fundus images medical dataset.Specifically:(1)An automatic glaucoma classification model based on deep convolutional neural network(DCNN)was constructed.On the one hand,due to the quality of fundus images and the specificity of characteristics of glaucoma,we designed a preprocessing method for glaucoma fundus image.Besides,we set up comparative experiments to verify the influence of these factors on DCNN network.On the other hand,we explored the advantages and disadvantages between adaptive extraction features(using deep learning network)and traditional artificial pre-defined features in the glaucoma classification task.Comparative experiments show that the features of the automatic extracted by DCNN are effective and bring accurate expression to the classification task.(2)An automatic glaucoma classification model based on transfer learning was constructed.The aim is to solve the contradiction between limited training samples and deep learning needs in a specific field.We used complete image dataset of cataract as the source domain to optimize the target domain(glaucoma)task.In addition,we integrated two domain features to further improve the generalization ability of the transfer learning model.Experimental results show that the proposed method improves the performance and efficiency of the original DCNN glaucoma automatic classification model.(3)A visual network of domain feature migration process was constructed.The interpretability of predictive models is significant in the medical field.However,deep learning is a "black box" model(where results are difficult to interpret).Based on our proposed transfer learning model,we can easily establish a Class Activation Mapping(CAM)module to explore the connection between the source domain of cataract and the target domain of glaucoma.The visualization results showed the changes of fundus characteristics during the whole transfer processing,which was in line with the clinical diagnostic criteria of ophthalmologists.