基于弱监督的改进型GoogLeNet在DR检测中的应用

针对糖尿病视网膜病变分级检测中标定样本少、多目标检测的问题,提出了一种基于改进型GoogLeNet的弱监督目标检测网络。首先,对GoogLeNet网络进行改进,去掉最后一个全连接层并保留检测目标的位置信息,添加全局最大池化层,以sigmoid交叉熵作为训练的目标函数以获得带有多种特征位置信息的特征图;然后,基于弱监督方法仅使用类别标签对网络进行训练;其次,设计一种连通区域算法来计算特征连通区域边界坐标集合;最后在待测图片中使用边界框定位病灶。实验结果表明,在小样本条件下,改进模型准确率达到了94.5%,与SSD算法相比,准确率提高了10%。改进模型实现了小样本条件下端到端的病变识别,同时该模型的高准确率保证了模型在眼底筛查中具有应用价值。

Application of improved GoogLeNet based on weak supervision in DR detection

To handle the issues of small sample size and multi-target detection in the hierarchical detection of diabetic retinopathy, a weakly supervised target detection network based on improved GoogLeNet was proposed. Firstly, the GoogLeNet network was improved, the last fully-connected layer of the network was removed and the position information of the detection target was retained. A global max pooling layer was added, and the sigmoid cross entropy was used as the objective function of training to obtain the feature map with multiple feature position information. Secondly, based on the weak supervision method, only the category label was used to train the network. Thirdly, a connected region algorithm was designed to calculate the boundary coordinate set of feature connected regions. Finally, the boundary box was used to locate the lesion in the image to be tested. Experimental results show that under the small sample condition, the accuracy of the improved model reaches 94%, which is improved by 10% compared with SSD (Single Shot mltibox Detector) algorithm. The improved model realizes end-to-end lesion recognition under small sample condition, and the high accuracy of the model ensures its application value in fundus screening.