| Glaucoma is the leading irreversible blindness disease in the world,and the new cases of glaucoma in China are increasing year by year with the aggravation of population aging.Glaucoma can be divided into open-angle and angle closure glaucoma according to the structure of the Anterior Chamber Angle(ACA).Early diagnosis and treatment are very important for reducing the blindness rate of Glaucoma.Optical coherence tomography(OCT)is widely used in glaucoma screening due to its non-invasion,non-intervention and quickness while gonioscopy is the gold standard for clinicians to assess ACA.Therefore,it is of great significance to study the effective classification of two types of images for the diagnosis of glaucoma based on the characteristics of ACA.The deep learning method provides an important means for the classification of ACA in the two kinds of images,but it has the following problems:(1)The speckle noise in OCT images and Gaussian blur in gonioscope images conceal key information,thereby influencing the prediction based on learning algorithms;(2)For the tri-classification of the ACA in OCT images,the single convolution-based models have a limited receptive field and the multiple convolution ones ignore the correlation of features at different scales,which makes it difficult to accurately extract the features when the Region of Interest(ROI)in the images changes greatly;(3)For the five-classification of the ACA in gonioscope images,the involved key structure is thin and long.The low-resolution models and reconstruction based high-resolution ones are easy to loss such information during feature conversion due to the limitation of network structure,which will affect the ACA classification.To address the above isses,this disseration has done the following work.Firstly,aiming at the speckle noise in the OCT image and the Gaussian blur in the gonioscope image,two corresponding image restoration algorithms are proposed,namely,the guided filtering based non-local means despeckling algorithm and a deep learning deblurring method.The former uses non-local information based guided filtering to extract feature information and gray information of the original noisy image and use them to estimate its similarity weight.Moreover,the boosting algorithm and iterative strategy are introduced to further improve the noise reduction performance.The method has achieved good performance on the retina and ACA in OCT images.The latter uses a coarse-to-fine multi-scale strategy to process the blurred image.A global information fusion and reconstruction network is proposed to integrate the multi-scale output features to further improve the global spatial information.The experiments performed on the natural images and gonioscope images have verified the feasibility and superiority of the algorithm.Secondly,because the ROI in the Anterior Segment-OCT varies greatly,a hybrid attention based pyramidal network is proposed to divide the ACA into open angle,narrow angle and closed angle.Pyramidal convolution includes different levels of kernels,where each scale includes filters with different depths and sizes.Thus,it can effectively capture different levels of subtle information and fully explore the correlation between different features through skip connection.Moreover,a hybrid attention module including spatial and channel attention is introduced to highlight important features.Experiments show that this method can provide better classification performance than the current mainstream networks in terms of accuracy,specificity,sensitivity on 422 test images,and its classification performance has been improved after image denoising.Finally,because the ROI in the gonioscope image is thin and long,a hybrid attention based densely connected high resolution network is proposed.Most of existing network models transform feature maps from high-resolution low level features to low-resolution high level ones.The proposed network will retain high-resolution feature maps,then gradually introduce low-resolution feature maps,and finally combine feature maps between different scales and depths to improve the high-level feature representation ability of different resolution images.With the hybrid attention module,the useful information will be strengthened to improve the accuracy.Experiments show that the algorithm can achieve96.18% accuracy and better results than the state-of-art models on 497 test dataset,and its classification performance has been improved on the deblurred images.Indeed,it provides an effective way to realize the automatic classification of gonioscope images. |
