融合注意力的ConvNeXt视网膜病变自动分级

融合注意力的ConvNeXt视网膜病变自动分级 Automatic classification of retinopathy with attention ConvNeXt 1 corresp first-author E-mail：huangwenbo@sina.com E-mail：huangwenbo@sina.com 黄文博（1980-），男，吉林长春人，博士，教授，硕士生导师，2018年于吉林大学获得博士学位，主要从事模式识别、图像处理及自然语言处理方面的研究。E-mail： huangwenbo@sina.com http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461846&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461852&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461850&type= 22.01333237 32.00399780 黄文博 （1980-），男，吉林长春人，博士，教授，硕士生导师，2018年于吉林大学获得博士学位，主要从事模式识别、图像处理及自然语言处理方面的研究。E-mail： huangwenbo@sina.com huangwenbo@sina.com 1 2 1 长春师范大学 计算机科学与技术学院，吉林 长春 130032 长春师范大学 计算机科学与技术学院，吉林 长春 130032 School of Computer Science and Technology， Changchun Normal University， Changchun 130032， China School of Computer Science and Technology， Changchun Normal University， Changchun 130032， China 中国科学院重大任务局，北京 100864 中国科学院重大任务局，北京 100864 Bureau of Major Tasks， Chinese Academy of Sciences， Beijing 100864， China Bureau of Major Tasks， Chinese Academy of Sciences， Beijing 100864， China 由于视网膜病变的类间图像特征差别小及分类临界值相对模糊，自动分级算法存在识别与分级准确率低的问题，提出了融合高效通道注意力（Efficient Channel Attention， ECA）特征的ConvNeXt视网膜病变自动分级模型。针对数据集中数据不足的问题，采用水平翻转左右变换的方法扩充数据，并引入相关数据集来均衡数据的分布。针对眼底图像中出现的图像模糊、光照不均等问题，采用Graham方法对图像进行预处理突出病变特征。提出了融合注意力的ConvNeXt网络来辅助医生诊断视网膜病变，引入ECA机制，并设计了E-Block模块，该模块具有高性能、低参数的特性，能够在训练过程中有效捕捉跨通道交互的信息，同时避免降维。采用迁移学习方法训练网络的所有层参数，加入dropout方法避免ConvNeXt网络的学习能力过强导致的过拟合问题。实验结果表明，所提出的模型敏感性为95.20%，特异度为98.80%，准确率为95.21%。与常用的网络相比，本文方法针对视网膜病变自动分级各项性能指标均有提高。 Due to the small differences in image features between classes and the relative fuzzy classification threshold of retinopathy， automatic classification algorithms are challenged by problems related to low recognition and classification accuracy. This paper proposes an automatic classification model for retinopathy based on an improved ConvNeXt network. Aiming at solving the problem of insufficient data in the data set， the horizontal flip left and right transformation method is used to expand the data， and related data sets are introduced to balance data distribution. To solve problems related to image blurring and uneven illumination in the fundus image， the Graham method was used to predict the image. The characteristics of the lesions are also highlighted. In this paper， an attention-fused ConvNeXt network was proposed to assist doctors in diagnosing retinopathy， an efficient channel attention mechanism was introduced， and an E-Block module was designed to channel interaction information while avoiding dimensionality reduction. The transfer learning method was used to train all layer parameters of the network， and the dropout method was added to avoid the overfitting problem caused by the strong learning ability of the ConvNeXt network. The results show that the sensitivity， specificity， and accuracy of the proposed model are 95.20%， 98.80%， and 95.21%， respectively. Compared with the ConvNeXt and other networks， the performance indexes of this network model for automatic classification of retinopathy. 视网膜病变识别与分级 迁移学习 ConvNeXt网络 高效通道注意力 E-Block retinopathy recognition and grading transfer learning ConvNeXt network Efficient Channel Attention（ECA） E-Block 1 引 言 1 视网膜病变会导致患者视物模糊、视力下降甚至失明 ［ 1 1 ］ ，由于缺少自动识别与分级系统，大量诊断工作主要由专家人工操作，因此基于深度学习方法的视网膜病变自动识别与分级成为研究热点 ［ 2 2 ］ 。 国内外学者对视网膜病变自动分级方法展开了较为深入的研究。Xu等 ［ 3 3 ］ 采用卷积神经网络（Convolutional Neural Network，CNN）将视网膜病变图像分类。Jiang等 ［ 4 4 ］ 通过预训练3个不同的CNN网络，通过Adaboost方法将预训练的模型融合实现视网膜病变识别。以上两种模型仅对患者是否存在视网膜病变进行判定，未得到具体的病变分级结果。目前，视网膜病变自动分级的方法较多，以CNN为基础模型进行改进或与其他方法相结合。Li等 ［ 5 5 ］ 提出了基于CNN的迁移学习方法，结合支持向量机（Support Vector Machine，SVM）实现视网膜病变分级。韦哲 ［ 6 6 ］ 提出利用基于迁移学习的Xception ［ 7 7 ］ 模型实现视网膜病变分级。王嘉良 ［ 8 8 ］ 提出了基于目标检测的区域全卷积网络（Region-based Fully Convolutional Networks，R-FCN）算法，通过检测血管瘤、眼底出血及玻璃体出血等病变区域，实现对视网膜病变的图像分级。Lin等 ［ 9 9 ］ 提出了抗噪声检测和注意力融合的视网膜病变检测模型。郑雯 ［ 10 10 ］ 引入空洞卷积代替普通卷积并融合注意力机制实现了视网膜病变分级。这些方法基本解决了视网膜病变分级问题中视盘区域对病灶识别的干扰问题，但由于视网膜微血管的类间图像特征差别小，分级的特征临界值相对模糊，现有视网膜病变分级算法仍存在分级准确率较低、模型能力受限等问题，因此目前视网膜病变分级仍然是一项具有挑战性的工作 ［ 11 11 ］ 。 本文提出一种基于ConvNeXt网络的视网膜病变自动识别与分级方法。ConvNeXt网络使用余弦下降法动态调整学习率。针对训练网络带标记的眼底图像数据不足，导致ConvNeXt网络在训练的过程中分级精度较低的问题，通过图像增强和引入额外数据集的方法解决数据分布不均衡问题。针对类间图像特征差别小导致的低精度问题，采用迁移学习法训练模型中所有层权重，设计了能够提取更深层次特征的E-Block模块，添加到ConvNeXt网络模型中，避免参数降维，有效捕捉跨通道交互信息，显著提升了模型的准确率。 2 图像预处理 1 2.1　数据集 2 本文的主体数据集采用APTOS 2019 Blindness Detection竞赛数据集。该数据集采用国际视网膜病变分级标准，依据视网膜病变的严重程度将图片分为5个等级，如 图1 图1 所示。其中，0级为正常，1级为轻度病变，2级为中度病变，3级为重度病变，4级为增殖性视网膜病变。通过观测数据集中各类数据的分布，可发现APTOS2019数据集存在严重的数据不均衡情况，导致模型更多关注样本多的类别，忽略样本少的类别，训练效果不理想。因此，这里引入由EyePACS平台提供的Kaggle（2015）竞赛数据集和IDRiD（Indian Diabetic Retinopathy Image Dataset）数据集。考虑到图像的采集环境，首先要对数据集进行清理，清理掉因受拍摄环境影响而出现的图像显示不正常和噪声太大的数据，如 图2 图2 所示。通过数据增强对数据量较少的类别（主要针对等级3和等级4）进行数据扩充，采用水平翻转、左右变换等方法扩充数据集，使得每个类别的图像数量相对均衡，最终构建成本文的实验数据集，共10 041幅图像，各类别图像数如 表1 表1 所示。 数据类别 扩充前 扩充后 0级-正常 1 805幅 2 019幅 1级-轻度病变 370幅 2 035幅 2级-中度病变 999幅 2 013幅 3级-重度病变 193幅 1 982幅 4级-增殖性DR 295幅 1 992幅 2.2　图像预处理 2 由于数据集中多数眼底图像的对比度比较低，病灶区域和非病灶区域不容易区分，使得病变特征不易检测。另外，受制于眼底图像的拍摄环境及成像设备的性能，数据集中的图像出现不同程度的图像模糊、照明不均等问题。为了得到更好的训练效果，本文使用Graham方法 ［ 12 12 ］ 对图像进行预处理，减轻因不同照明条件或成像设备而产生的问题。给定一张眼底图像 <math id="M1"><mi>I</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461886&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461882&type= 1.60866666 2.79399991 ，经Graham方法处理后的图像 <math id="M2"><mover accent="true"><mi>I</mi><mo>̂</mo></mover></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461892&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461889&type= 1.60866666 3.72533321 为： <math display="block" id="M3"><mover accent="true"><mi>I</mi><mo>̂</mo></mover><mo>=</mo><mi>α</mi><mi>I</mi><mo>+</mo><mi>β</mi><mi>G</mi><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo><mi>*</mi><mi>I</mi><mo>+</mo><mi>γ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461898&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461896&type= 34.45933533 4.57200003 ， 其中： <math id="M4"><mi>G</mi><mo stretchy="false">(</mo><mi>θ</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461904&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461901&type= 8.29733276 3.55599999 是具有标准差 <math id="M5"><mi>θ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461910&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461908&type= 1.86266661 2.79399991 的2D高斯滤波器，*是卷积算子， <math id="M6"><mi>α</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461916&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461914&type= 1.94733346 2.79399991 ， <math id="M7"><mi>β</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461924&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35461921&type= 2.28600001 3.64066648 和 <math id="M8"><mi>γ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462570&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462567&type= 1.94733346 3.64066648 是加权因子。 采用中心裁剪法去掉图像中含有无用信息的黑色区域，如 图3 图3 所示。处理后的图像裁剪掉了无关区域，提亮了眼球边缘的暗部区域，出血点、絮状物、渗出物和血管等重要特征更加突出。在训练时还引入了随机裁剪、随机透视变换和高斯滤波等方法对图像进行特征增强。 3 模型设计 1 3.1　网络搭建 2 视网膜病变图像分类的难点主要是不同类别图像的病变特征差距很小且重叠率高，属于细粒度图像分类的范畴，是对大类下的子类进行分类，只有向神经网络输送大量的图像信息才能挖掘到更深层次的特征。本文搭建的ConvNeXt网络结构如 图4 图4 所示，详细参数信息如 表2 表2 所示。选择ConvNeXt-T网络作为基础网络，与Transfrom网络 ［ 13 13 ］ 相比，ConvNext网络不需要进行分块合并、窗口偏移和相对位置偏执等操作，具有更好的性能和更少的计算量。ConvNext网络的整体结构与ResNet ［ 14 14 ］ 类似，在Block的堆叠次数设计上参考了Transformer网络。本网络主要由卷积层、E-Block层和线性层3部分组成。其中，E-Block层是本网络的核心部分，负责提取视网膜图像的特征，E-Block层1至E-Block层3由E-Block模块和下采样模块组成，E-Block层4由E-Block模块和全局平均池化层组成，E-Block层1至E-Block层4的堆叠次数分别为（3，3，9，3），延续使用ConvNeXt-T网络结构堆叠次数设置。 结 构 输 入 卷积核及步距 输 出 4 卷积层1 E-Block层1 224 <math id="M9"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 224 <math id="M10"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 3 4 <math id="M11"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 4， s4 56 <math id="M12"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 56 <math id="M13"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 96 56 <math id="M14"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 56 <math id="M15"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 96 d7 <math id="M16"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7， s1 56 <math id="M17"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 56 <math id="M18"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 96 56 <math id="M19"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 56 <math id="M20"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 96 1 <math id="M21"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1， s1 56 <math id="M22"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 56 <math id="M23"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 384 56 <math id="M24"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 56 <math id="M25"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 384 1 <math id="M26"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1， s1 56 <math id="M27"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 56 <math id="M28"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 96 4 下采样 E-Block层2 56 <math id="M29"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 56 <math id="M30"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 96 2 <math id="M31"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 2， s2 28 <math id="M32"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 28 <math id="M33"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 192 28 <math id="M34"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 28 <math id="M35"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 192 d7 <math id="M36"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7，s1 28 <math id="M37"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 28 <math id="M38"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 192 28 <math id="M39"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 28 <math id="M40"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 192 1 <math id="M41"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1，s1 28 <math id="M42"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 28 <math id="M43"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 768 28 <math id="M44"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 28 <math id="M45"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 768 1 <math id="M46"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1， s1 28 <math id="M47"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 28 <math id="M48"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 192 4 下采样 E-Block层3 28 <math id="M49"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 28 <math id="M50"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 192 2 <math id="M51"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 2， s2 14 <math id="M52"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 14 <math id="M53"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 384 14 <math id="M54"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 14 <math id="M55"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 384 d7 <math id="M56"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7，s1 14 <math id="M57"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 14 <math id="M58"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 384 14 <math id="M59"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 14 <math id="M60"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 384 1 <math id="M61"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1， s1 14 <math id="M62"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 14 <math id="M63"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1 536 14 <math id="M64"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 14 <math id="M65"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1 536 1 <math id="M66"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1， s1 14 <math id="M67"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 14 <math id="M68"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 384 4 下采样 E-Block层4 14 <math id="M69"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 14 <math id="M70"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 384 2 <math id="M71"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 2， s2 7 <math id="M72"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7 <math id="M73"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 768 7 <math id="M74"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7 <math id="M75"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 768 d7 <math id="M76"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7， s1 7 <math id="M77"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7 <math id="M78"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 768 7 <math id="M79"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7 <math id="M80"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 768 1 <math id="M81"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1， s1 7 <math id="M82"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7 <math id="M83"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 3 072 7 <math id="M84"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7 <math id="M85"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 3 072 1 <math id="M86"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 1， s1 7 <math id="M87"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 7 <math id="M88"><mo>×</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462533&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462530&type= 2.20133328 2.37066650 768 在ConvNeXt模型中，首先将经过预处理的视网膜图像输入模型，在第一部分对图像使用4×4的卷积核进行卷积处理，设置步距为4，经过层归一化（Layer Normalization，LN）处理后进入第二部分的E-Block层，最后在第三部分的线性层输出模型的预测结果。 3.2　E-Block模块 2 理论上模型的网络深度越深，其性能就会越好，但是大量实验表明，盲目地加深网络的深度会适得其反，造成“退化”现象。针对这个问题，何凯明等 ［ 14 14 ］ 在ResNet网络的设计中提出了残差单元结构，残差单元将经过处理的参数与原参数相加，解决了网络的退化问题。本文基于残差基本思想提出了E - Block结构。E - Block结构设计如 图5 图5 所示，相比于经典卷积方法，深度可分离卷积（Depthwise Separable Convolution，DSC）可在保留大卷积核的基础上减少参数量并加速模型训练。因此，本文首先利用DSC减少参数，加速模型训练。在DSC中，每个卷积核的通道数都等于1，每个卷积核只负责输入特征矩阵的一个通道数，所以卷积核的个数必须等于输入特征矩阵的通道数，从而使得输出特征矩阵的通道数等于输入特征矩阵的通道数；然后，通过两个1×1的卷积核来调整输出特征矩阵的通道数，再将调整后的特征输入到高效通道注意力模块（Efficient Channel Attention， ECA），进一步提取深层次特征；最后，为了防止模型出现过拟合现象，提高模型的泛化能力，设置了一个Dropout层和一个DropPath层。Dropout法通过随机删除神经元的方式来减少神经元之间的协同性，它强迫一个神经单元和随机挑选出来的其他神经单元共同工作。DropPath层的作用与Dropout层类似，不同的是DropPath法会将模型中的主分支结构按几率失活，即DropPath以一定的概率将主结构的输出变为0，等价于只有捷径分支构成输出。 3.3　高效通道注意力模块 2 ECA模块通过不降维的局部跨信道交互策略和自适应选择一维卷积来实现。该模块可在只引入少量参数的情况下获得明显的性能提升，如 图6 图6 所示。通过全局平均池化，将每个通道上对应的空间信息（ H * W ）压缩到对应通道中变为1个具体数值，此时一个像素表示一个通道，最终变为1×1× c ，形成一个向量。随后，通过一个卷积核尺寸为 k 的一维卷积对该向量进行特征提取，得到权重信息。在获得权重信息的过程中，通过参数 k 来控制第0~ k 个通道参与计算，这样就得到了第一个通道的权值。卷积核向下滑动1格，控制第1~ k +1个通道参与计算，这样便实现了局部跨通道交互，再通过卷积核滑动依次算出所有通道的权重信息。最后，将计算出的权重信息与对应通道的特征像素相乘就完成了特征图的重校验。为了避免在交叉验证时手动调整参数 k ，ECA模块以自适应方式来确定 k 值，见式（2）~ 式（4） 式（4） ，其中，交互的覆盖范围（卷积核尺寸为 k ）与通道维度成正比。 假设在 k 和 C 之间存在映射 <math id="M89"><mi>ϕ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462557&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462554&type= 2.20133328 3.64066648 ： <math display="block" id="M90"><mi>C</mi><mo>=</mo><mi>ϕ</mi><mo stretchy="false">(</mo><mi>k</mi><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462561&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462559&type= 14.81666565 3.97933316 . 由 <math id="M91"><mi>ϕ</mi><mo stretchy="false">(</mo><mi>k</mi><mo stretchy="false">)</mo><mo>=</mo><mi>γ</mi><mi mathvariant="normal">*</mi><mi>k</mi><mo>-</mo><mi>b</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462574&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462572&type= 24.04533195 3.97933316 可知，其中 <math id="M92"><mi>γ</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462570&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462567&type= 1.94733346 3.64066648 =2， b =1，线性函数是最简单的映射方式，但是以线性函数为特征的关系非常有限。由于通道维数 C 通常是2的指数倍，因此通过将线性函数 <math id="M93"><mi>ϕ</mi><mo stretchy="false">(</mo><mi>k</mi><mo stretchy="false">)</mo><mo>=</mo><mi>γ</mi><mi mathvariant="normal">*</mi><mi>k</mi><mo>-</mo><mi>b</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462574&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462572&type= 24.04533195 3.97933316 拓展到非线性函数来解决这个问题： <math display="block" id="M94"><mi>C</mi><mo>=</mo><mi>ϕ</mi><mo stretchy="false">(</mo><mi>k</mi><mo stretchy="false">)</mo><mo>=</mo><msup><mrow><mn mathvariant="normal">2</mn></mrow><mrow><mo stretchy="false">(</mo><mi>γ</mi><mi mathvariant="normal">*</mi><mi>k</mi><mo>-</mo><mi>b</mi><mo stretchy="false">)</mo></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462578&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462576&type= 29.29466629 4.14866638 . 给定通道维数 C 后可得： <math display="block" id="M95"><mi>k</mi><mo>=</mo><mi>ψ</mi><mo stretchy="false">(</mo><mi>C</mi><mo stretchy="false">)</mo><mo>=</mo><mo stretchy="false">|</mo><mfrac><mrow><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><msub><mrow><mi mathvariant="normal">g</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub><mo stretchy="false">(</mo><mi>C</mi><mo stretchy="false">)</mo></mrow><mrow><mi>γ</mi></mrow></mfrac><mo>+</mo><mfrac><mrow><mi>b</mi></mrow><mrow><mi>γ</mi></mrow></mfrac><msub><mrow><mo stretchy="false">|</mo></mrow><mrow><mi mathvariant="normal">o</mi><mi mathvariant="normal">d</mi><mi mathvariant="normal">d</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462581&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462579&type= 45.29666519 9.05933285 ， 其中 <math id="M96"><mo stretchy="false">|</mo><mfrac><mrow><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><msub><mrow><mi mathvariant="normal">g</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub><mo stretchy="false">(</mo><mi>C</mi><mo stretchy="false">)</mo></mrow><mrow><mi>γ</mi></mrow></mfrac><mo>+</mo><mfrac><mrow><mi>b</mi></mrow><mrow><mi>γ</mi></mrow></mfrac><msub><mrow><mo stretchy="false">|</mo></mrow><mrow><mi mathvariant="normal">o</mi><mi mathvariant="normal">d</mi><mi mathvariant="normal">d</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462586&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462583&type= 25.99266624 9.05933285 表示最接近 <math id="M97"><mfrac><mrow><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><msub><mrow><mi mathvariant="normal">g</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msub><mo stretchy="false">(</mo><mi>C</mi><mo stretchy="false">)</mo></mrow><mrow><mi>γ</mi></mrow></mfrac><mo>+</mo><mfrac><mrow><mi>b</mi></mrow><mrow><mi>γ</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462592&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462589&type= 20.57399940 9.05933285 的奇数。 4 实验结果与分析 1 实验在Pytorch深度学习框架下搭建，实验环境为python 3.8，pytorch 1.10，cuda 10.2。把数据集中，80%的图像划分为训练集，20%的图像划分为测试集。使用自定义学习率LambdaLR=6×10 -4 ，warmup=10，在达到最大学习率后采用余弦下降法动态调整学习率，优化器为AdamW，训练epoch=200，batch_size=64。本文使用迁移学习法，将ConvNeXt网络在ImageNet_1k上训练好的权重作为本文网络模型的初始权重。 4.1　评价标准 2 本文模型利用灵敏度（Sensitivity）、特异度（Specificity）和准确率（Accuracy）这3个指标来评估性能，主要由TP（True Positive）、FP（False Negative）、FN（False Negative）和TN（True Negative）4个基础指标计算，即： <math display="block" id="M98"><mi mathvariant="normal">S</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">n</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">v</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">y</mi><mo>=</mo><mfrac><mrow><mi>T</mi><mi>P</mi></mrow><mrow><mi>T</mi><mi>P</mi><mo>+</mo><mi>F</mi><mi>N</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462608&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462605&type= 36.66066742 8.04333401 ， <math display="block" id="M99"><mi mathvariant="normal">S</mi><mi mathvariant="normal">p</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">f</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">t</mi><mi mathvariant="normal">y</mi><mo>=</mo><mfrac><mrow><mi>T</mi><mi>N</mi></mrow><mrow><mi>T</mi><mi>N</mi><mo>+</mo><mi>F</mi><mi>P</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462614&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462610&type= 36.49133301 8.04333401 ， <math display="block" id="M100"><mi mathvariant="normal">A</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">u</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">y</mi><mo>=</mo><mfrac><mrow><mi>T</mi><mi>P</mi><mo>+</mo><mi>T</mi><mi>N</mi></mrow><mrow><mi>T</mi><mi>P</mi><mo>+</mo><mi>T</mi><mi>N</mi><mo>+</mo><mi>F</mi><mi>P</mi><mo>+</mo><mi>F</mi><mi>N</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462620&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=35462617&type= 54.86399841 8.12800026 . 由 表3 表3 可见，使用ConvNeXt网络训练扩充后的数据集相比扩充前各项指标均有提高，由此证明了本文数据均衡方法的有效性。使用本文方法训练Densenet网络和Shufflenet网络均取得了较好的分类结果，证明本文的训练策略具有优秀的鲁棒性。由于CBAM（Convolutional Block Attention Module）机制有助于网络提取更深层次的特征信息，所以本文在ConvNeXt网络中引入了CBAM注意力模块。实验结果表明，CBAM机制的引入与原ConvNeXt网络的表现几乎没有差别，其原因是引入CBAM的同时增加了更多参数，网络更加复杂，计算量增大，结果不如预期。本文还比较了SE（Squeeze - and - Excitation）通道注意力对网络的影响。实验结果表明，加入ECA模块的网络兼具通道注意力参数少、轻量化的优点，性能更好。改进的ConvNeXt网络的3项评价指标都在95%以上，对比其他算法，各项指标均有显著的提高且各项指标之间的参数差较小，说明本文设计的网络性能更稳定。 算 法 灵敏度 特异度 准确率 ConvNeXt（数据扩充前） 75.82 96.06 85.11 ConvNeXt 94.26 98.56 94.27 Densenet ［ 15 15 ］ 91.72 97.94 91.72 Shufflenet ［ 16 16 ］ 91.78 97.96 91.77 ConvNeXt+CBAM 94.18 98.54 94.16 ConvNeXt+SE 94.52 98.64 94.51 ConvNeXt+ECA 95.20 98.80 95.21 % 5 结 论 1 本文提出了一种结合E-block模块的改进ConvNeXt视网膜病变自动分级模型。利用迁移学习方法在训练时提高了模型训练速度，加快了模型收敛；改进模型的残差结构并设计具有高性能、低参数特性的E-block模块，避免参数降维，有效捕捉跨通道交互信息，在只增加少量参数的情况下获取到更多的图像细节，提升了模型准确率。实验结果表明，本文模型对视网膜病变的识别与分级敏感性为95.20%、特异度为98.80%，准确率为95.21%，各项指标均有显著提高，可为临床提供借鉴。 参考文献 LI T ， BO W ， HU C Y ， et al . Applications of deep learning in fundus images： a review ［J］. Medical Image Analysis ， 2021 ， 69 ： 101971 . doi: 10.1016/j.media.2021.101971 http://dx.doi.org/10.1016/j.media.2021.101971 Applications of deep learning in fundus images： a review ALYOUBI W L ， SHALASH W M ， ABULKHAIR M F . Diabetic retinopathy detection through deep learning techniques： a review ［J］. Informatics in Medicine Unlocked ， 2020 ， 20 ： 100377 . doi: 10.1016/j.imu.2020.100377 http://dx.doi.org/10.1016/j.imu.2020.100377 Diabetic retinopathy detection through deep learning techniques： a review XU K L ， FENG D W ， MI H B . Deep convolutional neural network-based early automated detection of diabetic retinopathy using fundus image ［J］. Molecules （Basel， Switzerland） ， 2017 ， 22 （ 12 ）： 2054 . doi: 10.3390/molecules22122054 http://dx.doi.org/10.3390/molecules22122054 Deep convolutional neural network-based early automated detection of diabetic retinopathy using fundus image JIANG H Y ， YANG K ， GAO M D ， et al . An interpretable ensemble deep learning model for diabetic retinopathy disease classification ［J］. Annual International Conference of the IEEE Engineering in Medicine and Biology Society IEEE Engineering in Medicine and Biology Society Annual International Conference ， 2019 ， 2019 ： 2045 - 2048 . doi: 10.1109/embc.2019.8857160 http://dx.doi.org/10.1109/embc.2019.8857160 An interpretable ensemble deep learning model for diabetic retinopathy disease classification LI X G ， PANG T T ， XIONG B ， et al . Convolutional neural networks based transfer learning for diabetic retinopathy fundus image classification ［C］. 10th International Congress on Image and Signal Processing， BioMedical Engineering and Informatics （CISP-BMEI）. October 14 - 16 ， 2017. Shanghai. IEEE ， 2017： 1 - 11 . Convolutional neural networks based transfer learning for diabetic retinopathy fundus image classification 韦哲 ， 赵刚 ， 王能才 ， 等 . 基于深度学习的糖尿病视网膜分类方法研究 ［J］. 中国医学装备 ， 2020 ， 17 （ 10 ）： 11 - 15 . doi: 10.3969/J.ISSN.1672-8270.2020.10.004 http://dx.doi.org/10.3969/J.ISSN.1672-8270.2020.10.004 基于深度学习的糖尿病视网膜分类方法研究 WEI ZH ， ZHAO G ， WANG N C ， et al . Study on the classification method of diabetic retina based on in-depth learning ［J］. China Medical Equipment ， 2020 ， 17 （ 10 ）： 11 - 15 . （in Chinese） . doi: 10.3969/J.ISSN.1672-8270.2020.10.004 http://dx.doi.org/10.3969/J.ISSN.1672-8270.2020.10.004 Study on the classification method of diabetic retina based on in-depth learning CHOLLET F . Xception： deep learning with depthwise separable convolutions ［C］. 2017 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. July 21 - 26 ， 2017 . Honolulu， HI. IEEE ， 2017 ： 1800 - 1807 . doi: 10.1109/cvpr.2017.195 http://dx.doi.org/10.1109/cvpr.2017.195 Xception： deep learning with depthwise separable convolutions 王嘉良 ， 罗健旭 ， 刘斌 ， 等 . 基于R-FCN算法的糖尿病眼底病变自动诊断 ［J］. 计算机工程与应用 ， 2020 ， 56 （ 4 ）： 109 - 114 . 基于R-FCN算法的糖尿病眼底病变自动诊断 WANG J L ， LUO J X ， LIU B ， et al . Automatic detection of diabetic retinopathy based on R-FCN algorithm ［J］. Computer Engineering and Applications ， 2020 ， 56 （ 4 ）： 109 - 114 . （in Chinese） Automatic detection of diabetic retinopathy based on R-FCN algorithm LIN Z W ， GUO R Q ， WANG Y J ， et al . A Framework for Identifying Diabetic Retinopathy Based on Anti-noise Detection and Attention-based Fusion ［M］. Medical Image Computing and Computer Assisted Intervention-MICCAI 2018 . Cham ： Springer International Publishing ， 2018 ： 74 - 82 . doi: 10.1007/978-3-030-00934-2_9 http://dx.doi.org/10.1007/978-3-030-00934-2_9 A Framework for Identifying Diabetic Retinopathy Based on Anti-noise Detection and Attention-based Fusion 郑雯 ， 沈琪浩 ， 任佳 . 基于Improved DR-Net算法的糖尿病视网膜病变识别与分级 ［J］. 光学学报 ， 2021 ， 41 （ 22 ）： 72 - 83 . doi: 10.3788/aos202141.2210002 http://dx.doi.org/10.3788/aos202141.2210002 基于Improved DR-Net算法的糖尿病视网膜病变识别与分级 ZHENG W ， SHEN Q H ， REN J . Recognition and classification of diabetic retinopathy based on improved DR-net algorithm ［J］. Acta Optica Sinica ， 2021 ， 41 （ 22 ）： 72 - 83 . （in Chinese） . doi: 10.3788/aos202141.2210002 http://dx.doi.org/10.3788/aos202141.2210002 Recognition and classification of diabetic retinopathy based on improved DR-net algorithm 谢云霞 ， 黄海于 ， 胡建斌 . 基于深度卷积神经网络的糖尿病视网膜病变分期及病灶检测 ［J］. 计算机应用 ， 2020 ， 40 （ 8 ）： 2460 - 2464 . doi: 10.11772/j.issn.1001-9081.2019122198 http://dx.doi.org/10.11772/j.issn.1001-9081.2019122198 基于深度卷积神经网络的糖尿病视网膜病变分期及病灶检测 XIE Y X ， HUANG H Y ， HU J B . Staging and lesion detection of diabetic retinopathy based on deep convolution neural network ［J］. Journal of Computer Applications ， 2020 ， 40 （ 8 ）： 2460 - 2464 . （in Chinese） . doi: 10.11772/j.issn.1001-9081.2019122198 http://dx.doi.org/10.11772/j.issn.1001-9081.2019122198 Staging and lesion detection of diabetic retinopathy based on deep convolution neural network GRAHAM B . Kaggle diabetic retinopathy detection competition report ［R］. University of Warwick， Tech . Rep.， 2015 . Kaggle diabetic retinopathy detection competition report LIU Z ， LIN Y T ， CAO Y ， et al . Swin transformer： hierarchical vision transformer using shifted windows ［C］. 2021 IEEE/CVF International Conference on Computer Vision （ICCV）. October 10 - 17 ， 2021 . Montreal， QC， Canada. IEEE ， 2021 ： 10012 - 10022 . doi: 10.1109/iccv48922.2021.00986 http://dx.doi.org/10.1109/iccv48922.2021.00986 Swin transformer： hierarchical vision transformer using shifted windows HE K M ， ZHANG X Y ， REN S Q ， et al . Deep residual learning for image recognition ［EB/OL］. IEEE ， 2016 ： 770 - 778 . 2015 ： arXiv ： 1512 . 03385 . https：//arxiv.org/abs/1512.03385 https://arxiv.org/abs/1512.03385 . doi: 10.1109/cvpr.2016.90 http://dx.doi.org/10.1109/cvpr.2016.90 Deep residual learning for image recognition HUANG G ， LIU Z ， VAN DER MAATEN L ， et al . Densely connected convolutional networks ［C］. 2017 IEEE Conference on Computer Vision and Pattern Recognition . Honolulu， HI， USA . IEEE ， 2016 ： 2261 - 2269 . doi: 10.1109/cvpr.2017.243 http://dx.doi.org/10.1109/cvpr.2017.243 Densely connected convolutional networks ZHANG X Y ， ZHOU X Y ， LIN M X ， et al . ShuffleNet： an extremely efficient convolutional neural network for mobile devices ［C］. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition . Salt Lake City， UT， USA . IEEE ， 2018 ： 6848 - 6856 . doi: 10.1109/cvpr.2018.00716 http://dx.doi.org/10.1109/cvpr.2018.00716 ShuffleNet： an extremely efficient convolutional neural network for mobile devices 10.37188/OPE.20223017.2147 TP391.7 A 1004-924X（2022）17-2147-08 黄文博,黄钰翔,姚远等.融合注意力的ConvNeXt视网膜病变自动分级[J].光学精密工程,2022,30(17):2147-2154. 黄文博,黄钰翔,姚远等.融合注意力的ConvNeXt视网膜病变自动分级[J].光学精密工程,2022,30(17):2147-2154. HUANG Wenbo,HUANG Yuxiang,YAO Yuan,et al.Automatic classification of retinopathy with attention ConvNeXt[J].Optics and Precision Engineering,2022,30(17):2147-2154. HUANG Wenbo,HUANG Yuxiang,YAO Yuan,et al.Automatic classification of retinopathy with attention ConvNeXt[J].Optics and Precision Engineering,2022,30(17):2147-2154. 吉林省自然科学基金资助项目（No.YDZJ202101ZYTS147）；吉林省教育厅科学研究规划项目（No.JJKH20200830KJ）；吉林省教育厅科学技术研究项目（No.JJKH20210887KJ） 2022-05-31 2022-06-17 2022-09-10 2022-10-08 光学精密工程 17 30