广西科技大学学报2025，Vol.36Issue(3)：92-99,8.DOI:10.16375/j.cnki.cn45-1395/t.2025.03.012
基于多尺度密集连接的糖尿病视网膜病变分类网络
The classification network for diabetic retinopathy based on multi-scale dense connections
摘要
Abstract
Diabetic retinopathy can be effectively prevented from worsening through early screening.The utilization of deep learning techniques in medical-assisted diagnosis improves the efficiency of diagnosis.In order to enhance the practical application of deep learning in medical-assisted diagnosis,this study focused on reducing the parameter size and complexity of network models.A lightweight backbone network for classifying diabetic retinopathy based on multi-scale dense connections was proposed.The feature-guided attention module was employed to enable the backbone network to focus on important regions within the lesions and extract discriminative and expressive lesion features.Additionally,a multi-scale feature dense module was designed to enable the network to capture both global and detailed information,thereby enhancing the network's perception of lesion features and compensating for the loss of lesion information during downsampling,ultimately improving feature extraction capability.Experimental results on the APTOS2019 dataset demonstrate an accuracy of 0.838 8 for the proposed network.Compared with similar networks,the method presented in this study achieves excellent performance with low parameter size and complexity.关键词
早期筛查/轻量化/多尺度特征/密集连接/分类网络Key words
early screening/lightweight/multi-scale features/dense connections/classification network分类
医药卫生引用本文复制引用
秦丽君,林川..基于多尺度密集连接的糖尿病视网膜病变分类网络[J].广西科技大学学报,2025,36(3):92-99,8.基金项目
国家自然科学基金项目(62266006,61866002) （62266006,61866002）
广西自然科学基金项目(2020GXNSFDA297006,2018GXNSFAA138122,2015GXNSFAA139293)资助 （2020GXNSFDA297006,2018GXNSFAA138122,2015GXNSFAA139293）
