Image Processing And Hierarchical Diagnosis Of Congenital Cataract Based On Semantic Segmentation

Congenital Cataract is a type of eye disease that often occurs around birth or in childhood,which seriously affects the visual development of children and has become the second leading cause of childhood blindness.According to statistics,22% to 30% of children’s blindness are caused by congenital cataract,but the method of manual diagnosis is time-consuming,laborious,and subjective,so the use of artificial intelligence technology to assist doctors in the diagnosis and treatment of congenital cataract patients has great clinical and social significance.Clinically,the main basis for the diagnosis of congenital cataract are slit-lamp images of the patient’s eyes taken by the slit-lamp microscope,and the most focused part is the lens region of slit-lamp images.However,actual slit-lamp images often contain irrelevant noises of eyelids,iris,eyelashes,etc.Furthermore,there are reflections of the slip-lamp microscope illumination source in the lens region.Hereinafter,the reflection of the slit-lamp source in the eyes is called the spot region,which interferes with the diagnosis of artificial intelligence models.In view of this,this paper studies the automatic segmentation of the lens and the spot region of the slit-lamp images based on the semantic segmentation method,and proposes a slit-lamp spot region restoration method based on the centroid position and a sliding window,which eliminates the interference of eyelids,spots and other noises to the diagnostic model.This paper finally proposes a multi-label classification network to achieve the hierarchical diagnosis of congenital cataract.The main work of this paper is as follows:(1)Research on the segmentation method of lens and spot region based on semantic segmentation.Since the captured slit-lamp images contain noises of eyelids,iris,eyelashes,etc.,it is necessary to filter out these interference noises by segmentation operation when analyzing them,and only retain images of the lens region as the input of the subsequent diagnosis model.This paper proposes an improved U-net network to achieve precise segmentation of the lens and spot region of slit-lamp images and compares it with the Deep Labv3+ segmentation model.The average pixel accuracy on the test set is up to97.57%,and the mean intersection over union reaches at 96.52%.(2)Research on the restoration method of the spot region in slit-lamp images.Since the illumination light source of the slit-lamp microscope will form several spots on the reflection of the eyes,these spots are similar in color to the cataract lesions,and most of the spots are in the lens region,which cannot be filtered out by the automatic segmentation operation of the lens,so the spot noises will inevitably interfere with disease diagnosis.This paper proposes a spot region restoration algorithm based on the position of the centroid position and a sliding window,which eliminates the interference of the spot inside the lens to the diagnostic model.Additionally,this paper uses the residual convolutional neural network to perform normal/diseased binary classification experiments on different data sets(original image,automatically segmented lens region image,lens region image after restoring light spots).As a result,the classification accuracy rates of the same network under the three data sets are 76.58%,93.36%,and 97.62%,respectively.This study shows that the segmentation and reduction operations have a great positive effect on the classifier.(3)Research on the classification and diagnosis of congenital cataract based on multi-label classification network.Clinically,there are three main classification criteria for the diagnosis of cataract patients: the area of the lesion region(large/small),the location of the lesion(yes/no completely covers the center of the visual field),and the compactness(deep/light).Based on this,this paper proposes a multi-label classification network for the classification and diagnosis of slit-lamp images.The output of the model is the classification of images under three classification standards,that is,three labels are output for each input sample,such as predicting a sample as a large lesion region,completely covering the center of the field of vision and having a deep compactness.The overall accuracy rate of the multi-label classification network reaches at 91.95%,which can provide ophthalmologists with more reliable auxiliary diagnoses.