摘要
针对糖尿病视网膜病变(DR)图像分类准确率不高的问题,提出了一种基于注意力机制的分类方法。该方法首先使用裁剪、均值滤波、高斯滤波、限制对比度自适应直方图均衡化(CLAHE)等方法进行预处理,模型以DenseNet架构作为基础模型,加入了通道注意力机制和空间注意力机制进一步提高模型对关键特征的识别能力。实验结果表明:与传统方法相比,改进后的模型分类效果优于传统模型,在Kaggle数据集其平均准确率达到了89.15%,平均特异度为94.22%;在超广角数据集其平均准确率达到了91.24%,平均特异度为95.72%。
To address the issue of low accuracy in image classification of Diabetic Retinopathy(DR),a classification method based on attention mechanisms is proposed.This method initially employs preprocessing techniques such as cropping,mean filtering,Gaussian filtering,and CLAHE,The model is based on the DenseNet architecture and incorporates channel attention mechanisms and spatial attention mechanisms to further enhance its ability to recognize key features.Experimental results show that,compared to traditional methods,the improved model exhibits superior classification performance.On the Kaggle dataset,its average accuracy reached 89.15%,with an average specificity of 94.22%.On the ultra-wide-angle dataset,its average accuracy reached 91.24%,with an average specificity of 95.72%.
作者
阚玉常
张光华
卓广平
汪扬
周金保
马非
KAN Yuchang;ZHANG Guanghua;ZHUO Guangping;WANG Yang;ZHOU Jinbao;MA Fei(School of Computer Science and Technology,Taiyuan Normal University,Jinzhong 030619,China;Department of Computer Science and Technology,Taiyuan University,Taiyuan 030032,China;Shanxi Eye Hospital,Taiyuan 030002,China)
出处
《太原学院学报（自然科学版）》
2024年第1期30-37,共8页
Journal of TaiYuan University:Natural Science Edition
基金
山西省科技攻关计划(202203021211006)
山西省科技攻关计划(YDZJSX2022B015)。
