基于微小病灶关注网络的糖尿病性视网膜病变自动诊断  

Automatic Diagnosis of Diabetic Retinopathy Based on Micro-lesions Attention Network

作　　者：向子健 宁春玉[1] 于洁茹 XIANG Zijian;NING Chunyu;YU Jieru(School of Life Science and Technology,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学生命科学技术学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2023年第4期136-142,共7页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：吉林省科技发展计划项目(20220101123JC)。

摘　　要：糖尿病性视网膜病变是一种致盲率极高的眼科疾病。临床中,医生须根据视网膜的病变程度对患者采取相应治疗措施。但由于患者众多、病灶微小,因此在诊断过程会出现误诊或漏诊现象。对此,结合视网膜中病灶的尺寸差异,提出了一种基于微小病灶关注的视网膜病变自动诊断模型。该模型以改进的ResNet34为主干网络,融入局部特征提取网络(LFENet)来增强对微小病灶的捕获能力,并采用迁移学习解决数据稀缺问题。模型在双网络的驱动下,实现病变诊断与病变程度的分级,在测试集上的准确率分别达到0.9701和0.7962,参数量仅有5.5 M,与其他网络模型相比,准确率更高、参数量更低,具有临床应用价值。Diabetes retinopathy is an eye disease with a high rate of blindness.In clinical practice,doctors treat patients according to the degree of retinal lesions.However,due to the large number of patients and small size of the lesions,misdiagnosis or missed diagnosis may occur in the diagnosis process.To address the above situation,combined with the size difference of the lesions in the retina,an automatic diagnosis model of retinopathy based on the focus of micro-lesions is proposed.The model uses improved ResNet34 as the backbone network,incorporates local feature extraction network(LFENet)to enhance the ability to capture small lesions,and uses transfer learning to solve the problem of data scarcity.Driven by the dual network,the model realizes the diagnosis and grading of pathological changes,and its accuracy on the test set reaches 0.9701 and 0.7962 respectively,with only 5.5 M parameters.Compared with other network models,the accuracy of this model is higher,and the number of parameters is less,which has clinical application value.

关 键 词：糖尿病性视网膜病变 微小病灶 迁移学习 ResNet34 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]