Research And Implementation Of Multi-label Classification Model For Fundus Medical Image Based On Graph Neural Network

Based on the existing research foundation in the field of medical image processing,it can be found that the development of deep learning technology in the recognition and processing of medical images has become relatively mature,especially in the research of disease classification and diagnosis tasks based on medical images.In the field of ophthalmic imaging alone,there are a large number of algorithms based on convolutional neural network(CNN)models that use color fundus imaging and OCT images to achieve results close to those of professional doctors in the recognition tasks of diabetic retinopathy(DR),diabetic macular edema(DME),glaucoma,and other diseases.However,there are still some problems with AI algorithms before they can be applied to the actual diagnosis and treatment of wet AMD in real-world scenarios.First,existing technology has not achieved fine-grained diagnosis of subtypes of wet AMD,which limits the application of the model in clinical diagnosis and treatment.Second,mainstream CNN models as black box methods completely abandon the prior knowledge of experts in the field.In contrast to the process of doctors identifying symptoms and inferring the type of disease the patient is suffering from based on evidence-based medicine,AI recognition methods lack recognition and inference of patient symptoms,as well as interpretability.Finally,the algorithm lacks matching system support and practicality.In response to the above-mentioned challenges,this paper conducts research in the following three areas:1)Enhanced attention-based fine-grained classification diagnosis model for wet AMD:This paper proposes the SAE-wAMD model,which is an improvement on the traditional VGG16 convolutional neural network model,utilizing self-attention layers to replace certain convolutional layers,thereby allowing the model to aggregate global features and improve the capture of critical lesion features.Additionally,the model utilizes contrastive learning loss functions for unsupervised pre-training.Experimental results demonstrate the effectiveness of the SAE-wAMD model,achieving an F-1 score of 86.81%.This represents a 4.4%improvement over the baseline model and outperforms the SOTA model,as well as surpassing the average level of four professional ophthalmologists.2)GCN based multi-label disease classification model for retinal medical images:This paper proposes the wAMD-GCN model,which first utilizes a convolutional backbone network to extract symptom features from retinal images.The corresponding category features are then embedded as nodes in the graph convolutional layer,leveraging the graph convolutional neural network’s feature aggregation capabilities to explore hidden correlations between diseases and symptoms,thus improving classification accuracy and interpretability.Experiments on a dataset of 4,927 accurately annotated OCT images demonstrate that the introduction of symptom labels improved the model’s F-1 accuracy by 3.7%.3)Development of a wet AMD diagnosis system:Based on the research results of the two models above,this system utilizes a wet AMD diagnosis database to build an intelligent diagnosis system that can assist doctors and alleviate the conflict between the workload of doctors and the demand for patient care.The platform includes disease assistance diagnosis,user and data management,and diagnostic report generation.This system overcomes the shortcomings of previous research by considering clinical practicality and offering services that are applicable to clinical scenarios.Ultimately,the models and system presented in this paper have been preliminarily verified through clinical applications at the Xingtai Eye Hospital in Hebei Province.These findings have laid a solid foundation for the development of multi-disease classification models for other retinal medical images and intelligent auxiliary diagnostic research for other types of retinal diseases.