| Dry eye refers to a disease that the quality or quantity of tears is abnormal for any reason,which damages the integrity of the tear film and causes eye discomfort.Meibomian gland dysfunction is a complex,chronic and common disease that affects the quality of tears in the eye.Accurate segmentation of the meibomian gland region is the basis for automatic measurement of morphological parameters and diagnosis of dry eye disease.However,due to the infrared imaging caused by artifacts and scattered noise,the variability of glandular regions in different meibomian glands and the discontinuity of glandular region boundaries,accurate segmentation of meibomian glands in meibomian gland infrared images becomes a very challenging task.In view of these problems,this thesis takes the data set of meibomian gland infrared image as the research object,and studies the segmentation and quantification index of meibomian gland based on eye image.The main work is as follows:Firstly,the data set of meibomian gland image segmentation is constructed,and the original data set image is preprocessed,including : image normalization and local Laplacian pyramid filtering method and hole filling to avoid the interference caused by uneven light,and high-quality labeling of the image.Secondly,the eyelid image is segmented based on Mask and Otsu threshold segmentation method,and a method based on statistical segmentation and gradient segmentation is used to segment the glands.The intersection ratio and accuracy of the healthy group and the dry eye group obtained by this method were greater than 70%,but it was not enough to support clinical application.Therefore,a new network segmentation model(ST-Uper Net)based on Swin Transformer and Uper Net is proposed for the segmentation of glandular regions.The model is improved based on the Uper Net structure,and the backbone network uses a lightweight Swin Transformer module for feature extraction,which reduces parameter calculation and memory requirements.In the feature extraction module of ST-Uper Net,the Co T attention module is added to filter the background information,increase the fusion of features,and improve the accuracy of feature extraction.Then,a point-based rendering Point Rend neural network is introduced to render the image segmentation more finely,so as to reduce the loss of glandular edge detail information in the image,and effectively reduce the loss of feature details and depth information.Compared with four classical segmentation networks,such as U-Net,experiments are carried out.The results show that the meibomian gland segmentation model based on ST-Uper Net is significantly better than the other four models in multiple evaluation indicators.Among them,the average accuracy and average intersection ratio reached 86.39%and 77.83%,respectively.The accuracy,recall rate and F1 reached 73.96%,73.30% and73.63%,respectively.The ablation experiment proves the effectiveness of the improved module in this thesis.Finally,the length,width,curvature,number,density and gland loss ratio of the segmented meibomian glands were calculated and analyzed.The results showed that the changes of gland morphology were related to dry eye: the decrease of length,number and density,the increase of loss ratio,the increase of width and the increase of curvature indicated that the gland deteriorated and was likely to form dry eye. |
PDF Full Text Request