摘要
目的:基于注意力机制与经典深度学习模型提出糖尿病眼底图像分类算法,提高糖尿病眼底图像识别的准确率。方法:使用包含5种类别共4 581张图像的开源数据集,进行特征提取、图像增强、批处理、洗牌等操作构建数据张量,将注意力机制与ResNet模型相结合构建ResNet-At模型,比较该模型与经典深度学习模型CNN、ResNet、AlexNet在识别糖尿病眼底图像方面的效果。结果:基于注意力机制的ResNet-At模型的准确率、精确率、召回率、F1值分别为89.5%、74.3%、62.4%、0.678,均高于其他经典深度学习模型。结论:基于注意力机制与经典深度学习模型的糖尿病眼底图像分类算法可以提高糖尿病眼底图像分类的识别效果。
Objective To propose a diabetic fundus image classification algorithm based on attention mechanism and classical deep learning model to improve the accuracy of diabetic fundus image recognition.Methods An open source dataset containing 5 categories of 4,581 images was used to perform feature extraction,image enhancement,batch processing,shuffling and other operations to construct the data tensor.The ResNet-AT model was construct by combining the attention mechanism with the ResNet model,which was compared with the classical deep learning models CNN,ResNet,AlexNet’s effect on the ecognition of diabetes fundus images.Results The accuracy,precision,recall and F1-score of the deep learning model based on the attention mechanism were 89.5%,74.3%,62.4% and 0.678,respectively,which were higher than other classical deep learning models.Conclusion The diabetic fundus image classification algorithm based on attention mechanism and classical deep learning model can improve the effectiveness of identifying diabetic fundus image classification.
作者
彭乔立
么冬爱
Peng Qiaoli;Yao Dong’ai(Information Center, Zhongnan Hospital of Wuhan University;Physical Examination Center, Zhongnan Hospital of Wuhan University)
出处
《中国数字医学》
2022年第11期58-61,共4页
China Digital Medicine
基金
武汉大学中南医院科技创新培育基金面上培育项目(CXPY2020040)。
关键词
糖尿病眼底图像
图像识别
深度学习
注意力机制
Diabetes fundus images
Image recognition
Deep learning
Attention mechanism
