Research On Automatic Diagnosis Of Pediatric Cataract Retro-illumination Image And Postoperative Complications Prediction Based On Pattern Recognition

Pediatric cataract is an ophthalmic disease seriously harming the visual development of children and has become one of the leading cause of blindness in children,which is a great threat to the health and life of patients.In recent years,as the wide application of slit-lamp image technology in ophthalmic disease diagnosis,slit-lamp examination has become an important means of pediatric cataract diagnosis.Therefore,computer aided diagnosis based on image recognition has formed an emerging pediatric cataract medicine mode.Besides,a large majority of patients will suffer from some unexpected postoperative complications like Posterior Capsular Opacification or High Intraocular Pressure within one year after the operation which replace the sick lens with an artificial one,which tends to interfere with postoperative recovery and visual reconstruction,even lead to blindness again,but the factors causing complications have not yet been figured out.Notwithstanding the above,since the group with frequent occurrence are mainly the elderly,and there exists few medical record for pediatric cataract samples,the related work on pediatric cataract is still in a preliminary stage.In this paper,the retro-illumination image data of pediatric cataract and the structured medical record data of postoperative complications we used comes from Zhongshan Ophthalmic Center,Sun Yat-sen University.Based on the retro-illumination images,we combined different visual feature extraction methods such as wavelet transform and LBP with several classifiers like SVM,k NN and LDA to construct predicting models.Together with the deep learning method DBN,we finally got 9 methods to make predictions on the 3 grading problem(area,density and location)of pediatric cataract,and then a performance comparison was made among different methods.The sensitivity of parameters in 9 methods was also analyzed in detail.Furthermore,we trained the naive Bayes classifier and random forest respectively with the structured medical record data of postoperative complications and made predictions on the occurrence of 2 major kinds of complications.Since the dataset is imbalanced,it has been processed by SMOTE oversampling method.Previous achievement of this work has been published in Nature Biomedical Engineering and reported by IEEE Spectrum.In this paper,for the grading diagnosis research based on retro-illumination images,a combined method using BOW model to make feature extraction and SVM to make classification achieved the highest accuracy over 95% with few misdiagnosis and missed diagnosis cases.Thus,the predicting model of pediatric cataract grading in this paper has shown an outstanding performance and high reliability.In the issue of postoperative complication prediction,except one case,both Naive Bayes and random forest classification models reached an accuracy superior to 80%.