淋巴结转移检测的八度卷积方法

针对乳腺癌淋巴结转移的人工检测精度低、耗时长的问题，提出了一种基于残差网络结构，使用八度卷积（Octave Convolution）方法设计卷积层的一种神经网络检测模型。以残差网络中卷积层为基础，首先对卷积层中输入和输出的特征向量进行高低分频，并缩减低频通道的宽和高为高频的一半。然后通过对减半后的低频向量进行上采样，实现与高频向量之间的卷积操作；通过对高频向量的平均池化，实现与减半低频向量的卷积操作。最后将高频向量之间、高频与低频向量之间的卷积相加，得到高频输出；将低频向量之间、低频与高频向量之间的卷积相加，得到低频输出。这样就构建出了Octave卷积层，将残差网络中所有卷积层完全替换为Octave卷积层就搭建出了检测模型。理论上，Octave卷积层中卷积的运算量减少了75%，有效加速了模型的训练。在最大内存为13 GB，可用磁盘大小为4.9 GB的云端服务器上，用PCam数据集进行测试，得到模型的识别精度为95.06%，内存占用8.7 GB，磁盘占用356.4 MB，平均单次训练用时4 min 42 s。与ResNet50相比，所提模型的精度下降了0.69%，内存节省0.6 GB，磁盘节约105.9 MB，单次训练用时缩短了1 min。实验结果表明，所提模型具有较高的识别精度，较短的训练用时和较少的内存消耗，使得模型在大数据的背景下，降低了对算力资源的要求，具有一定的实际应用价值。

Octave convolution method for lymph node metastases detection

Focused on the problems of low accuracy and long time cost of manual detection of breast cancer lymph node metastasis, a neural network detection model based on residual network structure and with Octave convolution method to design convolution layers was proposed. Firstly, based on the convolution layer of residual network, the input and output eigenvectors in the convolution layer were divided into high frequencies and low frequencies, and the channel width and height of the low-frequencies were reduced to half of those of the high frequencies. Then, the convolution operation between the low-frequency vector and the high-frequency vector was realized by up-sampling the low-frequency vector with the reduction by half, and the convolution operation between the high-frequency vector and the low-frequency vector was realized by average pooling of the high-frequency vector. Finally, the convolutions between high-frequency vectors and between high-frequency vector and low-frequency vector were added to obtain the high-frequency output, and the convolutions between low-frequency vectors and between low-frequency vector and high-frequency vector were added to obtain the low-frequency output. In this way, Octave convolution layer was constructed, and all convolution layers in residual network were replaced by Octave convolution layers to construct the detection model. In theory, the amount of computation of convolution in Octave convolution layer was reduced by 75%, effectively speeding up the training of the model. On the cloud server with maximum memory of 13 GB and free disk size of 4.9 GB, the PCam (PatchCamelyon) dataset was used for testing. The results show that the model has the recognition accuracy of 95.1%, the memory occupied of 8.7 GB, the disk occupied of 356.4 MB, and the average single training time of 4 minutes 42 seconds. Compared with the ResNet50, this model has the accuracy reduced by 0.6%, the memory saved by 0.6 GB, the disk saved by 105.9 MB, and the single training time shortened by 1 minute. The experimental results demonstrate that the proposed model has high recognition accuracy, short training time and small memory consumption, which reduces the requirement of computing resources under the background of big data era, making the model have application value.