Research On Intelligent Aided Diagnosis Systems For Primary Angle Closure Glaucoma

Glaucoma,as the world’s first irreversible blinding eye disease,is a serious threat to human visual health.Primary angle closure glaucoma(PACG)is the main type of glaucoma in China.PACG has no obvious clinical symptoms at the early stage,and it is generally found when the disease develops to the middle and late stage and the visual field is seriously deficient,which seriously damages the visual health of our people.At present,the total amount of ophthalmic medical resources in China is insufficient,and the distribution of high-quality medical resources is uneven,which can not fully meet the clinical needs.The lack of ophthalmic medical resources makes PACG screening in China face severe challenges.With the aging of China’s population,the situation of glaucoma prevention and treatment in China is becoming more serious.computer-aided diagnosis systems such as deep learning can quickly complete a large number of repetitive medical activities,which is expected to reform ophthalmic diagnosis and treatment from the supply-side,and is of great significance to the prevention and treatment of glaucoma in China.Ultrasound biomocroscopy(UBM)can present the shape of the anterior chamber angle visually,which has a unique advantage in the diagnosis of angle-closure glaucoma.Under the new situation of national strategies such as Healthy China and artificial intelligence,this study carried out research on computer-aided diagnosis systems for PACG based on deep learning algorithm and UBM image,expecting to provide auxiliary reference information for the screening,diagnosis and treatment of PACG.The UBM image dataset is firstly established.High-quality clinical data sets are the basis for the study of computer-aided diagnosis systems.Therefore,UBM images of patients were collected from real clinical environment in this study.The ophthalmologists who collected the images all had rich clinical experience and follow relatively consistent operation procedures.After desensitization and strict access criteria,a total of 2146 open angle images and 1642 angle closure images were included in the study.The detection of angle closure is an important basis for the diagnosis of PACG.In this thesis,neural networks such as ResNet50,InceptionV3 and InceptionResNetV2 are selected as feature extractors,and a customized classifier was designed for the angle closure recognition task.The automatic recognition of a angle closure was realized by combining technologies of data augmentation,transfer learning,and feature fusion.For unclosed images,the quantitative measurement of anterior chamber angle is an important reference for the risk assessment of angle closure.In this thesis,a localization network for anatomical key points was firstly designed based on EfficientNet network,and the automatic localization of scleral spur and angle recess was realized.Then,a semantic segmentation network was designed based on the atrous convolution module and EfficientNet to realize automatic segmentation of anterior chamber angle structure contour.According to the definition,the trabecular-iris angle(TIA),angle-opening distance(AOD)and angle recess area(ARA)were quantitatively calculated.The results show that the accuracy,sensitivity and specificity of ResNet50 and InceptionResNetV2 fusion model are 98.69%,97.60%and 99.53%,respectively,with good recognition performance.The mean errors of the key point localization model for scleral spur and angle recess were 65.19±51.47 μm and 43.32 ± 41.23 μm,respectively,and the mean errors of intraobserver localization were 62.42 ± 49.04 μm and 41.26±38.86μm,respectively.It shows that the performance of the model is close to the precision of ophthalmologist.The average IoU and Dice coefficient of the segmentation model were 97.11%and 98.53%,respectively.Compared with manual measurements,the coefficients of variation of TIA500,TIA750,AOD500,AOD750,ARA500 and ARA750 measured by deep learning system were 5.77%,4.67%,10.76%,7.71%,16.77%and 12.70%,respectively.And the reproducibility were 5.77 degrees,4.56 degrees,155.92μm,147.51 μm,0.10 mm2 and 0.12 mm2.The ICC values(intra-group correlation coefficients)of all the angle parameters were more than 0.935.It shows that the angle parameters measured by the deep learning system are in good agreement with the measured values by artificial.Subtype classification of potential mechanism of angle closure is helpful to guide clinical treatment.In order to identify the mechanism of angle closure of angle closure image and narrow angle image,neural networks such as InceptionV3 were selected as feature extractors,and custom classifiers were designed for this task.Then,automatic recognition of potential mechanism of angle closure was realized by combined data augmentation,transfer learning and class weight technology.The exact match ratio and hamming loss of the model are 65.93%and 0.1392 respectively,which verifies the feasibility of deep learning to identify the potential mechanism of angle closure.Based on deep learning models designed above,computer-aided diagnosis systems for PACG is designed in this thesis.Based on the UBM image,the system can quickly realize the intelligent recognition of the angle clouse,the intelligent quantitative evaluation of the angle parameters and the intelligent recognition of the potential mechanism of angle closure.The results are helpful for the early screening,diagnosis and management of PACG and has a good clinical application prospect.