一种基于双流CNN模型的糖尿病视网膜病变检测新方法

摘要

视力损害的主要成因——尤其是糖尿病视网膜病变（DR）与年龄相关性黄斑变性（AMD）——给临床诊断与治疗带来了严峻挑战。早期发现与及时干预有助于避免患者出现严重后果。本研究提出一种利用双流卷积神经网络（CNN）模型检测眼部疾病的新方法：第一条支路处理经预处理的眼底图像，第二条支路则在空间频率域中分析经高通滤波的眼底图像。为评估模型性能，本文采用APTOS 2019数据集——该数据集最初为2019年亚太远程眼科学会失明检测竞赛而整理，并在Kaggle上公开。本文方法的准确率达0.986，有望成为DR检测的早期筛查工具。

Abstract

Major causes of visual impairment, particularly diabetic retinopathy (DR) and age-related macular degeneration (AMD), have posed significant challenges for clinical diagnosis and treatment. Early detection and prompt intervention can help prevent severe consequences for patients. The study presents a novel approach for detecting eye diseases using a two-stream convolutional neural network (CNN) model. The first stream processes pre-processed fundus images, while the second stream analyzes high-pass filtered fundus images in the spatial frequency domain. To assess the model's performance, we use the APTOS 2019 dataset, which was originally compiled for the Asia Pacific Tele-Ophthalmology Society 2019 Blindness Detection competition and is publicly available on Kaggle. Our method shows promise as an early screening tool for DR detection with an accuracy of 0.986.