Research On The Key Problems In Automatic Diagnosis Of Ophthalmic Diseases Based On Machine Learning

Eyes,which have a close relationship with the central nervous system,are the important refractive structure in the body of human beings.80%-90%external information is passed to the brain via eyes,variety ophthalmic diseases are undoubtedly the main threaten for this precise refractive structure.There are researches showing that ophthalmic diseases severely affect the development of the vision of human beings.According to the WHO(World Health Organization),ophthalmic diseases,including ametropia,have become the third greatest threaten for the health and life quality of human after tumors,cardiovascular and cerebrovascular diseases.In our country,up to five million blind people account for 18%of the world’s total.Then,the distribution of the healthcare resource is unbalanced in our country,the doctors in high level hospitals must spent more time to tackling common diseases.Machine learning can alleviate this condition and improve the health circumstance in our country.This topic selects several key problems in the path of diagnosis and treatment of ophthalmology to talk about——automatic evaluation of the visual acuity of children,automatic diagnosis of multiple ophthalmic diseases and prediction of the postoperative complications of pediatric cataracts.Because the definition of above three problems is definitely,and much repeated work and diagnosis which occupies much time can be done by machine learning which is well applicable without delay.Difficulties in the application of machine learning in these problems——the amount of samples is small,the distribution of data is imbalanced,interpretability is not robust,so many emergency situations in data and poor clinical applicability——were detailed discussed.Then methods for solving these problems were proposed.(1)For the imbalanced distribution of samples,unexplainable and multiple emergence situations of data in the automatic evaluation of visual acuity of children.First,a method which makes use of the experience from doctors to build human-in-the-loop deep learning system is proposed.This system adopted object localization to detect main objects in the frames of video.In entire architecture,the division between binocular examination and monocular examination,whether the level of the visual acuity of child is uncooperative,N/A,larger than or equal to 0.21 is judged with the localization results of Faster-RCNN.Finally whether the level of the visual acuity of the child is[0.01-0.05]or[0.08-0.15]is classified with convolutional neural network.In this whole process,several groups of thresholds,which are set to judge some categories,were optimized with 88 short videos.The final accuracy is 75.54%in the last testing stage.In the discussion between doctors and computer system,the evaluation of doctors and computer systems own their merits,and the experiences from doctors could be more and more accurate during this form of discussion.Finally,human-in-the-loop deep learning system come into being.For the multiple thresholds of human-in-the-loop deep learning method,an presentation of video with seven channels including all information about different eye position,direction of visual acuity card,the level of visual acuity card and a 1-dimension convolutional neural network based evaluation method was proposed,and the ROI pooling layer was inserted into 1 dimension convolutional neural network to make the length of this presentation(input of fully connected layer)consistent.Since this classification problem is imbalanced,convolutional neural network with class weights is used.Finally,the classification accuracy is improved a little compared with the first method.In the contest with three clinical doctors,both of these two methods beat human.(2)For the unexplainability and imbalanced distribution of samples,Visible Genome——unified interpretable and expandable automatic diagnosis framework of multiple ophthalmic diseases based on deep learning was proposed.This framework diagnoses multiple ophthalmic disease with a unified thought and mode,and could provide diagnosing mind and reason that are similar to doctors’.The first stage of this framework is primary diagnosis of diseases with an accuracy of over 93%,including cataracts,pterygium,keratitis,subconjunctival hemorrhage and normal.The second stage is localization of anatomical parts and main foci,the localization results are over 82%and 92%for the slit lamp images under natural light and cobalt blue light.The third stage focuses on the attributes determination of anatomical parts and main foci,convolutional neural networks with class weights were used to solve 10 classification problems with accuracies of 79%-98%.Whereas,the classification performance with original images are almost same to the classification with anatomical parts and main foci,which show that convolutional neural network can understand which parts makes this image was categorized to this class in the classification problems with a bigger targets.The fourth stage is providing treatment recommendation according to the queries of the state of illness and the information obtained in previous three stages,and the necessity for performing surgery for pterygium patients is solved with convolutional neural network.A web based artificial intelligence(AI)platform is developed and deployed in the clinical practical of department of AI and big data,Sun Yat-sen University Zhongshan Ophthalmic Center to help patients.(3)For the small amount of cases,imbalanced distribution of samples,whole framework for predicting the postoperative complications of pediatric cataracts patients was proposed.Firstly,the minimum support and confidence of apriori algorithm was set as 0 to mine discriminant association rules,which can provide reference to doctors,among the dataset of not suffering complications and the dataset of suffering complications.The discriminability of association rules in real clinical scenes could be freely adjust by the discriminant threshold.Then three fold cross validation is used to testify the performance of na?ve Bayesian classifier and random forest in predicting the postoperative complications.SMOTE algorithm was used to enrich the amount of positive samples,whether suffers from complications,whether suffers from severe lens proliferation into the visual axis and abnormal high intraocular pressure were predicted respectively.Both the accuracies for two methods are over 74%,high false negative rate obtained with the original dataset is alleviated by SMOTE.Finally,genetic feature selection was used to choose the attributes which have closer relationship with complications.Experimental results show that complications未患病relationship with gender and the age at surgery;abnormal high intraocular pressure未患病relationship with secondary IOL placement,operation mode,age at surgery and area of cataracts attributes;severe lens proliferation into the visual axis未患病relationship with gender,operation mode and laterality.In the additional external testing,the accuracies three problems are over 66%.A web based online prediction platform was developed to help the works of doctors and show the association rules.Although this topic make many progresses,there are still some drawbacks.Firstly,prediction of the postoperative complications of pediatric cataracts,automatic diagnosis of multiple ophthalmic diseases need more clinical data to validate,and more clinical attributes can be made use of to improve the accuracy.Beside,automatic evaluation of visual acuity of children can be deployed in a wearable computing device and need to be modified with a real time judgement of examination result,eventually a system without human interference will come into being.