Research On The Classification Of Anterior Chamber Angle Images Based On Deep Learning

Glaucoma is a leading cause of irreversible blindness in the world.The main basis for determining clinical treatment protocols when glaucoma is diagnosed is evaluating the Anterior Chamber Angle（ACA）level with gonioscopy by ophthalmologists.There are five ACA levels that correspond to different glaucoma treatments.To evaluate the ACA level,the ophthalmologist need examine the four local structures in ACA images: Schwalbe Line（SL）,Trabecular Meshwork（TM）,Scleral Spur（SS）,and Ciliary Body Band（CBB）,With the increasing number of glaucoma patients in recent years,the ACA evaluation meets the following problems:（1）The workload of ophthalmologists is high,and the diagnosis and treatment is low efficiency;（2）The availability of experienced ophthalmologists is severely sparse and unevenly distributed which cause many glaucoma patient not to be diagnosed in time.（3）The evaluation is highly dependent on ophthalmologists’ personal experience and status,and the results could be unstable and inconsistent.In recent years,Deep Learning（DL）technology has been widely used in medical image analysis and achieved much success.However,directly applying traditional DL models for ACA level classifications is suboptimal because of the following challenges:（1）The four structures concentrate in a small area of a full ACA picture and it is difficult to distinguish the structures from each other;（2）It is difficult to integrate the domain knowledge of text-modal data with that of ACA images;（3）Lack of training data on ACA image evaluation;To overcome the above challenges,we propose a deep neural network with weakly-supervised metric learning,attribute learning and attention mechanism.The main contributions of this paper include:（1）We propose an ACA image classification method named GCNET based on weakly-supervised metric learning.In GCNET,the auxiliary segmentation task is designed to enhance the ability to learn and represent local important features by segmenting SL,TM,SS,and CBB from the background;an embedding module based on metric learning is designed to improve the ability to distinguish the four structures;and an interactive communication module based on context aggregation is designed to realize knowledge sharing among different modules.Experiment results show that the classification accuracy of the GCNET is79.19%,which outperforms the reference models such as VGG,Goog Le Net,Res Net,Fix Match,CCT,and UPS.（2）Considering to the importance of domain knowledge in text-modal data,we propose a vision-text fusion-based ACA image classification method named VTFNET.In VTFNET,two branches are designed to jointly learn vision and text information based on weakly-supervised metric learning and attribute learning;attention mechanisms and channel-wise concatenation are used to improve the correlation among multi-modal data and to achieve feature fusion.Experiment results show that the classification accuracy of the VTFNET is 81.83%,which outperforms GCNET.（3）A dataset ACA999 is constructed for ACA classification,which contains 999image-level labels,100 pixel-level labels.