Pathological Myopia Classification And Chorioretinal Atrophy Segmentation Based On Fundus Images

In recent years,with the steady increase of myopia and high myopia patients and the growing younger age trend,myopia has become a public health problem worldwide.What is more,pathological myopia is one of the main causes of visual impairment and even blindness.Chorioretinal atrophy is an important myopic retinopathy,which is also an important basis for clinical diagnosis and classification of pathological myopia.Therefore,the classification of pathological myopia and segmentation of atrophy based on fundus image are significant for the prevention,early diagnosis and treatment of pathological myopia.In this thesis,an improved residual network based on ResNet and attention mechanism is proposed to realize the automatic classification of pathological/non-pathological myopia.The attention mechanism is implemented by Optimized Global Context(OGC)block,which is first proposed in this thesis.First,expected maximization algorithm is introduced to improve channel attention module.And then the optimized channel attention module integrates the Global Context(GC)block to achieve global context capture.The OGC block is applied before the classifier layer of the proposed classification network,which can effectively improve the feature selection and presentation ability of the proposed network.An improved U-Net based method is proposed to realize the automatic segmentation of chorioretinal atrophy in fundus image.Residual structure,deep supervision mechanism and attention mechanism are introduced to improve the performance of the U shape segmentation network.An auxiliary loss function based Euclidean distance transformation(EDT)is proposed.The residual structure comes from the layer structure of ResNet34,which improves the depth of the network and enhances the feature extraction capability of the network.Inspired by M-Net,we introduce deep supervision mechanism to strengthen the constraint on the early layers during training,which can improve the accuracy of feature learning.The OGC block based on attention mechanism is applied in the decoder path to improve the ability of senior semantic feature selection.The EDT auxiliary loss function introduces spatial information for overall loss function of the segmentation network and improves the smoothness of segmentation results.The proposed pathological myopia classification and segmentation method was validated and evaluated on the public dataset,which is provided by ISBI 2019 Pathological Myopia Analysis Challenge.The dataset consists of 400 training images and 400 test images.(1)For the pathological/non-pathological myopia classification,the AUC value of the proposed method on test set is 0.9992,which shows that the proposed OGC attention module can effectively improve the feature learning ability of the classification network.(2)For the segmentation of chorioretinal atrophy,due to the great difference between pathological and non-pathological myopia eyes,the atrophy segmentation network was trained based on the pathological myopia classification results respectively.The average Dice coefficient of the proposed segmentation network is 82.20%.The experimental results show that the improved network structure and the proposed loss function are effective.