Recognition And Localization Of Pediatric Cataract Lesions Based On Target Detection

Pediatric cataract is a common ophthalmological disease which seriously endangers the vision healthy development of infant.It usually occurs during the sensitive period of visual development of infants.If not timely and effective medical diagnosis and treatment,it will cause irreversible visual impairment.Clinically,diffuse light source images are the main medium for ophthalmologists to diagnose pediatric cataract.However,pediatric cataracts are diverse and it is difficult to directly give detailed lesion descriptions.Therefore,the accurate lesion analysis is undoubtedly the key to establish early screening,diagnosis and monitoring management of pediatric cataract.With the development of artificial intelligence,classification and target detection algorithms based on deep learning have been widely used in the medical field.Computer aided automatic diagnosis opens up a new research direction for medical images analysis.The accumulation of diffusion light source images also provides data support for the automatic detection of pediatric cataract lesions,making it possible to locate and identify pediatric cataract lesions objectively,quickly and accurately based on computer algorithms.Based on the pediatric cataract diffuse light source images provided by a famous eye hospital in China,this paper constructs a pediatric cataract lesions automatic detection model based on the deep learning target detection algorithm.And based on this model,carrying out multi-angle lesions grading recognition to achieve the accurate analysis of pediatric cataract lesions.The research content mainly includes the automatic detection and grading of pediatric cataract lesions.In lesion detection,the noise reduction algorithm based on pixel four neighborhood relation is used to preprocess the pediatric cataract diffuse light source images and filter the noise.Then,target detection algorithm Faster R-CNN with a skip connection structure is proposed,which fully integrated the shallow location information and the deep semantic information,so as to conduct a research of automatic detection on lesions and improve the detection accuracy effectively.The study on lesion grading mainly includes three angles analysis of the lesion opacity density,area and location.The opacity density can be directly obtained by the detection model.The threshold calculation is carried out on the output of the detection model to obtain the grading of the opacity area and location of the lesion.Aiming at the uncertainty of threshold and lesion complexity,this paper proposes lesion grading model based on convolutional neural network and feature fusion.The lesion grading based on convolutional neural network is directly graded by deep convolutional neural network training model;The feature fusion of the lesions first extracts the primary features such as color,texture,shape,etc.,and then performs a variety of feature fusion.Then,the principal component analysis is used to reduce the dimension of the merged features,and finally,the classification is performed using Soft Max.Both models obtained more objective lesion grading results.The experimental results show that the lesion detection model with a skip connection structure achieved the optimal performance.The detection accuracy of the normal,lens,dense and transparent reached 99.8%,99.7%,90.4% and 89.4%,respectively,and has the hightest Mean Io U(0.9174).The Res Net-50 deep convolutional neural network obtained the best performance in the two-angle grading problem of the opacity area and position of the lesion,and the accuracy rates are 92.59% and 91.78%,respectively.The lesion detection and grading study in this paper provides a feasible scheme for the clinical application of computer aided diagnosis of pediatric cataract,and also provides a reference for other medical imaging studies.