Research On The Diagnosis Algorithm Of Keratitis Image Based On Deep Learning

Keratitis is one of the most common eye diseases causing blindness,which can cause very serious impact on the vision of patients,and even cause blindness in severe cases.At present,the clinical diagnosis of keratitis mainly rely on various ocular imaging examinations,but this approach is largely limited by the experience and expertise of the physician.At present,the research on automatic diagnosis of keratitis image mainly focuses on some specific medical images,some of which are difficult to take,some need special processing,and the process is more complex.Moreover,most of the studies only provide the image recognition results,but do not give the specific situation of keratitis.In view of this problem,this paper chose the diffuse light image of the anterior segment as the research object,which is easier to obtain and not only does no harm to the patients,but also allows the pathological changes and pathological process of keratitis to be visible through the photography of the anterior segment,which is convenient for disease tracking and case consultation.At the same time,this paper simulates the working process of doctors and designs a deep learning-based diagnosis algorithm for anterior segment images based on the interpretability of anterior segment images.First,the Res Net50,Inception V3 and Dense Net121 networks are used to classify normal corneal images and keratitis images.The transfer learning method is used to train the network model,and the Dropout regularization method is combined to prevent overfitting.And the models are analyzed and compared comprehensively from multiple indicators.Then based on the interpretability of keratitis images,the specific signs of keratitis are analyzed,and the diagnostic basis of keratitis is given,so as to facilitate doctors to understand the patient’s condition more quickly and assist doctors to complete the preliminary diagnosis of keratitis.Under the guidance of ophthalmologists,all anterior segment images are labeled from five signs,consisting of whether there is corneal opacity in the cornea,whether the boundary of the lesion is distinct,whether the epithelium in the lesion area is intact,whether there is congestion,and whether there is new blood neovascularization,which are important in the diagnosis of keratitis.Therefore,a multi-label image data set is constructed,and the data is enhanced by means of horizontal inversion according to the image characteristics.On the basis of the data set,an improved multi-attribute network based on Res Net50 network is constructed,including feature extraction part and classification part.Feature extraction part is used to extract image features,and classification part is a multi-output network with each channel corresponding to each attribute.In addition,the network model is trained by multi-task joint training,the feature extraction layers are shared among multiple tasks,and the correlation between different attributes is utilized to improve the accuracy of each attribute.The multiattribute network can recognize the keratitis image from five attributes at the same time,and finally get the specific characteristic pathological changes of keratitis,which can provide doctors with auxiliary diagnosis.The experimental analysis and comparison on the data set of the anterior segment show that the algorithm proposed in this paper has the advantages of high accuracy and fast recognition speed,and has certain practicability and expansibility,which is helpful for the auxiliary diagnosis of keratitis.