光学相干断层扫描技术是临床中检测视网膜病变的主要手段，但人工诊断的模式存在主观性强、效率低的问题，为此提出一种轻量级卷积神经网络用于视网膜病变的自动检测。所提网络由两种模块组成，第1种模块将空洞卷积与深度可分离卷积相结合以降低参数量；第2种模块利用分解卷积方法，通过将常规卷积层分解成多层不对称卷积的方式延展深度。两种模块交叉组合构成特征提取器，使用Softmax函数作为分类器，获得了44层深、参数量为9.2 MB的轻量级模型。所提网络在测试集上的准确率、敏感性、特异性、接收者操作特征曲线下面积分别达到0.980、0.954、0.987和0.997。可视化结果表明，模型诊断依据与眼科专家相一致。这些结果表明，所提网络能够准确地实现视网膜疾病的自动检测。

One major method for detecting retinopathy in clinics is optical coherence tomography. However, this manual diagnostic model is affected by strong subjectivity and low efficiency. Therefore, this paper proposes a lightweight convolutional neural network for the automatic detection of retinopathy. The proposed network consists of two modules. The first module combines atrous convolutions and depth wise separable convolutions to reduce the number of parameters; the second module uses the decomposition convolution method to extend the depth by decomposing the conventional convolution layer into multilayer asymmetric convolution. Both modules are combined to form a feature extractor, and the Softmax function is used as the classifier to obtain a lightweight model with 44 layers deep and 9.2 MB parameters. The accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve of the proposed network on the test set are 0.980, 0.954, 0.987, and 0.997, respectively. The visualization results show that the diagnostic basis of the model is consistent with that of ophthalmologists. These results show that the proposed network can accurately automate retinal disease detection.