Research On Choroidal Neovascularization Typing Method Based On Deep Learning

Age-related macular degeneration(AMD)is one of the leading causes of blindness,especially in the elderly.One of the characteristics of AMD is the presence of choroidal neovascularization(CNV).CNV refers to the new growth of blood vessels between the retina and choroid.After neovascular leakage and rupture,the patient’s macular is severely damaged after recurrence,which may lead to permanent visual impairment.Optical coherence tomography(OCT)is usually used for the diagnosis of CNV.OCT has the advantage of non-invasive and rapid acquisition of retinal images,so it can quantitatively and qualitatively evaluate CNV.OCT has become the first choice for clinical monitoring of CNV diseases.CNV classification is very important for the clinical treatment of CNV.In clinical treatment,CNV can be divided into type 1,type 2 and mixed type.At present,intravitreal injection of anti-vascular endothelial growth factor can effectively inhibit the development of CNV disease.However,the efficacy of different types of CNV injection of anti-vascular endothelial growth factor is different.The characteristics of high similarity between CNV classes and large intra-class differences make the classification task of CNV face challenges.Aiming at the problem of high similarity between CNV types,this paper achieves CNV typing by exploring subtle discriminative features between CNV types.The main methods include: learning local detail information and enhancing the semantic discriminative feature detail representation through feature fusion.Besides,aiming at the problem of large intra-class difference of CNV,this paper improves the performance of CNV typing by learning the discriminative features and intra-class diversity features of CNV.The main contributions are given as follows:(1)An erasing and refining discriminative feature for CNV typing is proposed.The network model is mainly composed of two parts: Finer feature mined(FIFM)and Attention-oriented feature fusion(AOFU).A new fine feature mining module is proposed,which forces the network to mine neglected local fine features and realize denoising and rich feature information.The attention-guided feature fusion module fuses the denoised salient features and local subtle features.The network performs deep learning with the help of the fusion feature enhancement stage to enhance the discrimination of the fusion features and improve the CNV typing performance.(2)A local detail enhancement network for CNV typing is proposed.The network model mainly includes two parts: progressive training and local detail enhancement(LDE)module.In the progressive training process,the features learned by the model combine shallow fine-grained information and high-level semantic information.In the LDE module,the detail feature learning technology is introduced to learn the underlying detail information.And then the detail information is embedded into the semantic feature map to enhance the high-level semantic detail information.Improving the network’s ability to learn the subtle differences between CNV classes.(3)A feature enhancement network for CNV typing is proposed.The network consists of two branches: discriminative feature enhancement branch and diversity feature enhancement branch.In the discriminative feature enhancement branch,a class-specific feature extraction(CSFE)module is introduced,in which the channel attention is used to learn discriminative features to improve the learning of CNV inter-class differences.In the diversity feature enhancement branch,Attention region selection(ARS)is used to mine diversity features from the feature map of the same CNV type to improve the learning of CNV intra-class diversity.