Research On Automatic Diagnosis And Prediction Algorithms For Pediatric Cataract Based On Slit-lamp Images

Pediatric cataract is a common ophthalmological blindness disease,accounting for about 40% of children blindness.Its occurrence and development seriously affect the physical health and quality of life of infants.If this disease is discovered and treated early,it can be completely cured.Therefore,the early diagnosis and prediction are the premise and key to the treatment of pediatric cataract.However,due to the great concealment of pediatric cataract,its manual diagnosis and prediction are very difficult,which leads to defects of time-consuming,laborious and subjective.On the other hand,with the development of artificial intelligence algorithms,big data technology and precision optical instruments,ophthalmic hospitals have accumulated a wealth of slit-lamp images,which provides a good research basis for the study of automatic diagnosis and prediction of pediatric cataract.However,the existing automatic diagnosis algorithms are still immature,and there are some limitations,such as relying heavily on doctors’ experiences,poor universality,low accuracy,high misdiagnosis and missed diagnosis rates,etc.Therefore,they cannot be effectively applied to clinical practice.In addition,pediatric cataract has its own characteristics such as high noises,complex phenotypes,severe heterogeneity of slit-images,limited sample,imbalanced data problem,unpredictable and lack of clinical application.Specialized studies on pediatric cataracts and their slit-lamp images are rare in existing literature.Hence,the automatic diagnosis and prediction of pediatric cataract is becoming one of the hotspots and difficulties in current artificial intelligence medical research.For the above problems,based on three kinds of slit-lamp images,this dissertation mainly studies the deep learning algorithms and their applications in preoperative screening,grading diagnosis,postoperative complications evaluation and progression trend prediction of pediatric cataract.(1).For the problems of large noises and complex phenotypes of the diffuse-illumination images,we propose a method(TCH-CNN)for automatic screening and grading diagnosis of pediatric cataract based on convolutional neural network(CNN).First,by analyzing the morphological information of the lens,the Canny operation and Hough transform are used to automatically locate the region of interest(ROI)of lens and filter out the surrounding noise.Second,combined with CNN,transfer learning and data augment methods,high-level features are extracted from the ROI of lens.Then,the SVM and softmax classifiers are applied to implement automatic screening and three-angles grading of pediatric cataract.The experimental results indicate that the features extracted from TCH-CNN are more distinguishable than those of four classical feature extraction methods(i.e.,local binary pattern,scale-invariant feature transform,wavelet transform,color and texture features).The average accuracy rates for screening and three-angles grading(opacity area,density,and location)are 97.07%,89.02%,92.68%,and 89.28%,respectively.Automatic localization and transfer learning methods have improved the accuracy of the TCH-CNN over the original CNN by 4.5%,9.07%,11.73%,and 9.66%,respectively.Based on the above results,a cloud-based computer-aided diagnosis software is developed and deployed,which has been used in the clinical diagnosis of Zhongshan Ophthalmic Center at Sun Yat-sen University and promoted to the grass-roots hospitals.The method and automatic diagnosis software proposed in this study have been reported by various media such as IEEE Spectrum and CCTV2.(2).For the severe heterogeneity,lower accuracy and sensitivity problems in grading diagnosis based on slit-illumination images,an ensemble learning of deep convolutional neural networks(DCNN-Ensemble)is proposed.After analyzing the complexity and heterogeneity of the slit-illumination images,combined with three convolutional neural networks with different structures,transfer learning and automatic localization of lens,the DCNN-Ensemble method greatly enhances the performance of grading diagnosis of pediatric cataract.The experimental results indicate that the DCNN-Ensemble algorithm is superior to the classical feature extraction methods,Easy Ensemble method and single convolutional neural network.The accuracy,specificity and sensitivity of three grading diagnosis are: opacity area(92.13%,92.00% and 92.31%),opacity density(92.77%,93.85% and 91.43%),and opacity location(92.76%,95.25% and 89.29%).Moreover,the DCNNEnsemble algorithm still maintains better stability on two external datasets.Finally,an automatic grading diagnosis software for pediatric cataract is developed and applied to the ophthalmology clinic.(3).For the postoperative posterior capsular opacification complication and imbalanced data problem of retro-illumination images,a cost-sensitive residual convolutional neural network(CS-Res CNN)algorithm is proposed.In this method,the cost-sensitive factor is integrated into the cross entropy loss function of the CS-Res CNN to distinguish the misclassification cost of different classes,and the grid-search procedure is employed to determine the appropriate range of cost-sensitive factor.The experimental results indicate that the CSRes CNN is far superior to the classical feature extraction methods and three data-level methods,and exhibits excellent performance in terms of accuracy,specificity and sensitivity(92.24%,93.19% and 89.66%).Compared with the native residual CNN,the sensitivity,F1-measure and G-mean of the CS-Res CNN increase by 13.61%,4.59% and 6.05%,respectively.Furthermore,combined with the data-level methods,the effectiveness and stability of the CS-Res CNN are verified from multiple perspectives.Based on this,an automatic evaluation software for postoperative complication is developed,applied and promoted in the clinic.(4).For the unpredictable problem of the progression of pediatric cataract,based on retroillumination sequence images derived from multiple re-examination stages,combined with transfer learning,automatic localization of lens,CNN and LSTM(long short term memory),an end-to-end temporal sequence prediction algorithm(Temp Seq-Net)is proposed.In this method,the CNN is employed to extract high-level features from retro-illumination images.Then,the features are fed into the LSTM to mine the relationship of sequence images and predict the progression of pediatric cataract on the next stage.From the aspects of effectiveness,efficiency and resource utilization,the performance differences between the six potential combinations of three CNNs(Alex Net,Goog Le Net and Res Net)and two sequence methods(LSTM and recurrent neural network)are compared.The experimental results indicate that the Alex Net-LSTM is optimal.The impact of sequence images with different lengths on the training and prediction are evaluated in detail.Only one model needs to be trained based on longer sequence data,which can simultaneously predict sequence images with lengths of 3–5.And the average time consumption of a single sequence is only 27.6ms,which achieves real-time prediction in clinic.This study provides an effective method for the automatic prediction of progression of pediatric cataract,which is beneficial for ophthalmologists to make treatment plan and alert patients in advance.Although the algorithms and system proposed in this dissertation have achieved some progress in the automatic screening,grading diagnosis,postoperative evaluation and prediction of pediatric cataract,they have only been validated on three different kinds of slitlamp images.The next step is to collect and arrange more medical images to verify the stability and universality of these algorithms.On the other hand,only a single type of slit-lamp images is used in the studies of the automatic diagnosis and prediction.The next step is to combine multi-source images and electronic medical records and design multimodal data processing algorithm to further reduce the misdiagnosis and missed diagnosios rates.In addition,the interpretability of the algorithms and the promotion of clinical applications are also the important issues for further research.