Research On Intelligent Diagnosis Of Fundus Lesions

In the context of the country's efforts to develop artificial intelligence and‘Healthy China',intelligent medicine with artificial intelligence as the core is regarded as an important development direction of future medicine,and it is also a multi-disciplinary high-tech technology.Accurate analysis of medical images is an indispensable technical means and tool for intelligent medicine in scientific research,clinical diagnosis,and treatment.In the field of ophthalmology,abundant imaging resources make it one of the most suitable research directions for artificial intelligence imaging diagnosis,showing broad clinical application prospects.In response to the ‘Thirteenth Five-Year Plan' of eye health planning,this work studies the deep convolutional neural network methods for the intelligent diagnosis of retinal fundus diseases,establishes neural network models for diabetic retinopathy identification and classification,neural network models for fundus screening,and neural network models for identifying abnormal parts of the fundus and diagnosing multiple fundus diseases.The corresponding intelligent auxiliary detection systems for fundus lesion examination are developed and put into trial use in some hospitals including Sichuan Provincial People's Hospital,showing good clinical applicability.The innovation points and main contributions of the thesis include:1.A method based on neural networks for identifying and grading diabetic retinopathy is proposed,which alleviates the problems of high imaging noise,limited medical data,and excessive temporal and spatial overhead of integrated models,obtains better prediction performance of models,effectively reduces the workload of ophthalmologist and improves the efficiency of the ophthalmologist.Diabetic retinopathy(DR)is the major cause of blindness in the working group,and it is also one of the serious complications of diabetes.There are more than 100 million diabetic patients in China,and the prevalence of DR is as high as 30%.The lack of ophthalmology resources and other reasons make DR screening face severe challenges.In response to this situation,this work studies a deep neural network method for DR intelligent screening and classification.Firstly,a multi-center DR data set is established,and a variety of image preprocessing methods are used to improve the image quality of fundus color photos and effectively alleviate the problem of excessive image noise.Then,aiming at the problem of limited data size in the research,using transfer learning technology,several two-stage deep neural network cascade learners are constructed.Secondly,because of the excessive time and space overhead in the training and storage of the integrated model,under the premise of reducing the training cost,to construct the optimal integrated model,the relationship between the ideal size of the ensemble model(the number of component classifiers)and the number of categories,and the relationship between the optimal width of the ensemble model(the combination method of component classifiers)and the number of categories are empirically discussed,respectively.Finally,a neural network model for DR screening and classification is proposed,and a DR intelligent screening and diagnosis assistant system is developed.The clinical trial results show that the system effectively reduces the workload of the ophthalmologist and improves the efficiency of the ophthalmologist.2.A method based on neural networks of fundus screening under Ultrawide-field(UWF)image quality enhancement is proposed,which alleviates the high missed diagnosis rate of the model caused by low image contrast,reduces the degree of model overfitting,and obtains good model prediction accuracy,and improves the feasibility of conducting fundus disease risk assessment in UWF images.The peripheral retina is a high incidence site for pathological changes of many fundus diseases.Traditional fundus color photographs cannot provide information about the periphery of the retina due to the limited imaging field of view,potentially leading to an increase in the missed diagnosis rate,and the examination process requires pupil dilation.In response to this problem,using the technical advantages of UWF fundus imaging for rapid,non-mydriatic,and panoramic imaging of the peripheral retina,this work establishes a multi-center UWF fundus image data set,and studies a deep neural network method for intelligent screening of fundus abnormalities and a deep neural network method for intelligent diagnosis of four blinding fundus diseases.The low contrast of the image leads to the high specificity and low sensitivity of the fundus disease screening neural network,and the limited disease data scale on this basis leads to serious overfitting of the diagnostic model.Therefore,a variety of image processing methods are used to improve UWF image quality,and then the effectiveness of different image quality enhancement methods to improve network prediction performance are specifically analyzed.Finally,a fundus intelligent screening model and a disease diagnosis model are developed by combining transfer learning technology and integrated learning technology.The rich experimental results show that these image processing methods can effectively improve the neural network model's ability to learn different pathological features of the fundus,so that it can obtain good prediction performance.The fundus intelligent screening model and the disease diagnosis model have potential application value for fundus health screening,and help improve the feasibility of clinicians in assessing the risk of ocular fundus lesions in patients.3.A method based on neural networks for detection of abnormal signs of the fundus and diagnosis of multiple diseases is proposed,which alleviates the problems of class imbalance and high similarity between classes,improves the diagnostic accuracy of the models for different fundus sign parts or diseases,and provides a more comprehensive and in-depth clinical reference information for eye health assessment.In this work,a multi-center large-scale UWF fundus image dataset is established,and a neural network model for early fundus automatic screening is developed.The diversity of retinopathy in the dataset is helpful to include heterogeneous populations in the fundus health screening.Then,because timely identification of fundus lesions located in the vitreous,retina,macula,and the optic nerve is of indispensable reference value for the diagnosis and treatment of many retinal diseases and systemic diseases,the UWF image data sets of these four fundus sites are further established.At the same time,because of the diversity of fundus diseases in the real scene and the urgent problem of referral,according to the relationship between the course of retinal tear and retinal detachment as well as their different treatment guidelines,the two diseases are further distinguished.A larger-scale UWF image fundus disease data set was established based on the previous work,combining the other two diseases(DR and high myopia pathological lesions).A deep neural network model for the recognition of abnormal signs of the fundus and a deep neural network model for the diagnosis of the above-mentioned diseases are developed.Besides,the problem of category imbalance and the similarity of features between categories is the difficulty of this research,and it is also the main reason for the high missed diagnosis rate and high misdiagnosis rate of the model.In response to these prob-lems,this work first adopts the category weight method to help the model automatically focus on samples from underrepresented categories.Then,a two-step classification strategy is proposed,so that the model can focus on learning the pathological characteristics of different fundus lesions or diseases while maintaining high specificity,thereby improving the diagnostic sensitivity of the model on each site or disease.Finally,a system for intelligent detection of abnormal signs of the fundus and intelligent diagnosis of the four diseases is developed.This system helps ophthalmologists to provide more comprehensive reference information for assessing the health of the fundus.