Deep Learning For Glaucoma Aided Diagnosis

Glaucoma is a group of blinding diseases characterized by progressive optic nerve damage that has become the leading cause of irreversible blindness in the world.Recent studies have shown that the prevalence of glaucoma will increase year by year as the aging population increases,and the number of glaucoma patients is expected to exceed100 million by 2040.The pathogenesis of glaucoma is not fully understood,and early glaucoma usually has no obvious symptoms and is insidious.This has resulted in a large number of glaucoma patients who are unaware of the illness and do not seek medical care until there is a noticeable visual impairment.Most cases are in the moderated and advanced stages of glaucoma at the first visit,so early screening and accurate diagnosis through ophthalmological examinations are essential to protect patients' vision.Ophthalmologists need to conduct a comprehensive analysis of various examination data such as structural and functional assessments to make an effective diagnosis,but this process is time-consuming,labor-intensive,and relatively subjective.It is difficult for the existing health workforce to achieve an ideal glaucoma screening effect.In recent years,deep learning has made great successes in computer vision and medical image processing.However,the glaucoma diagnosis algorithm based on deep learning technology has not received enough attention,especially the lack of research on multimodal deep learning diagnostic algorithms based on visual field(VF)and optical coherence tomography(OCT).In view of the urgent need for glaucoma automatic diagnosis tools in clinical practice,this dissertation conducts research on glaucoma aided diagnosis algorithms with deep learning.The main contents and innovations are as follows:(1)The first work collected more than 10,000 visual field reports from 7 ophthalmic centers in China to evaluate the feasibility of diagnosing glaucoma based on VF alone.In this work,we proposed a glaucoma diagnosis algorithm that integrates multiple statistical analysis plots of visual field and developed a glaucoma intelligent screening platform based on the algorithm.This work has three advantages as follows.First,the study sample size is large and there are many sources.Second,the performance of the model is improved by using multiple visual field analysis plots,and the algorithm achieves diagnostic performance superior to that of general ophthalmologists in tertiary hospitals.Finally,a software solution was developed to deploy the algorithm clinically.(2)Due to the limited information of VF to assess glaucoma,in the second work,we drew on clinical diagnosis experience and proposed a combined diagnosis algorithm based on VF and OCT.The complementarity of functional and structural assessments is exploited through an attention mechanism to improve diagnostic accuracy,and the domain generalization technique is introduced to improve the generality of the model.The algorithm is experimentally validated on data from multiple ophthalmic centers and multiple OCT devices.The experimental results show that the multimodal approach based on VF and OCT significantly improves the performance of the glaucoma diagnosis algorithm.(3)Medical studies have shown that there is a spatial correspondence between structural and functional impairments in glaucoma.However,the previous work neither fully utilizes the spatial correspondence between structure and function,nor does it have a multimodal interaction mechanism to realize the complementarity of the two modalities.The innovation of the third work is a deep learning based cross-modal relation reasoning network,which mainly contains relation reasoning and interaction transformer.The relation reasoning includes data-driven global relation reasoning and medical prior guided regional relation reasoning.The interaction transformer module designs a multimodal feature interaction technique to achieve global and regional structure-function feature fusion.Extensive experiments showed that VF and OCT data complement each other in identifying glaucoma,and this algorithm further improves the performance of the multimodal glaucoma diagnosis algorithm.(4)The multimodal algorithms proposed in the first two works require paired VF-OCT data to diagnose glaucoma.However,in clinical practice,collecting VF data is usually difficult,which may hinder the application of multimodal algorithms in screening scenarios.To obtain a more flexible clinical screening algorithm,the fourth work proposes a method to build a more accurate OCT model using OCT-VF data.In this work,we designed a generalized distillation framework for transferring the knowledge of the multimodal network(teacher)to the OCT network(student).Moreover,we proposed a novel asynchronous feature regularization module,which enables the student network to obtain more expressive OCT feature with the aid of VF information.In comparison with the regular trained OCT model,the OCT model obtained by this algorithm significantly improves the diagnostic performance of the single-modality glaucoma diagnosis algorithm and is comparable to the multimodal model.