多层次可选择核卷积用于视网膜图像分类

为了提高深度学习网络对糖尿病性视网膜病变识别准确率，针对光学相干断层扫描技术（optical coherence tomography，OCT）的视网膜图像分类研究，提出了一种基于可选择卷积核的网络模型，该模型能对多个尺度扩张率的卷积核之间进行自动选择操作。分割阶段生成多条路径，这些路径具有相同的卷积核但不同的扩张率，对应不同的神经元感受野大小；融合阶段将多条路径的信息进行组合和聚合，得到一个全局的、全面的选择权重表示；选择操作再根据2种权值自身相似性和相对相似性来选择权值。实验结果表明，该模型在2个视网膜公开的基准数据集OCT2017及SD-OCT上分别达到了95.39%，99.18%的分类结果。与目前已有的主流模型相比，该模型的实验结果在2个数据集上均有提升。

Multilevel selective kernel convolution for retina image classification

The paper presents a multilevel selective kernel convolution network (MSKNet) to improve the recognition accuracy of the deep learning network for diabetic retinopathy image classification of optical coherence tomography (OCT). First of all, the split stage generates multiple paths. These paths have the same convolution kernel but different dilation, corresponding to different receptive fields (RF). In the next part, multiple paths are combined and aggregated to obtain a global and comprehensive selection weight representation in the fusion stage. In the end, the selection operation selects the weights according to the feature similarity and relative similarity. The experimental results show that the proposed model achieves 95.39% and 99.18% classification results on the two open benchmark datasets OCT2017 and SD-OCT. Thus, compared with the existing mainstream models, the proposed model experimental results have advantages in both datasets.