Abstract:

Metastasis of breast cancer cells is an important factor affecting the prognosis of patients. The traditional pathologist’s examination process is redundant and time-consuming and easy to miss the micrometastases. At present, there have been achievements in the study of sentinel lymph node metastasis of breast cancer using convolutional neural network, but the accuracy is not high and the detection effect of micrometastasis is not good. Based on the breast cancer sentinel lymph node pathological image data set（PCam）, this paper designs and proposes a SENet multi-channel convolutional neural network model, which uses stacked multi-channel convolutional units and SENet modules, skipping cross-layer connections, standard convolution and depthwise separable convolution fusion, addition and concatenation operations. The model weight is obtained by using 50% image iteration training for 35 times, and then the accuracy and AUC value indexes are used to test the test image. The accuracy is 97.32% and the AUC value is 98.05%. Compared with the existing research results and the mainstream convolutional network model, the AUC value of this model ranks the first in the case of 49%, 51% and 100% test sets. The results show that this model has a high accuracy in the detection of lymph node metastasis, and a good detection performance for micrometastasis.

Key words: breast cancer, sentinel lymph node metastasis, PCam, multi-channel convolutional network, SENet

摘要：

乳腺癌细胞转移是影响病患预后的重要因素，检查前哨淋巴结病理切片可诊断转移情况。传统病理学医生检查过程冗余费时且极易漏检微转移病灶。目前已有利用卷积神经网络研究乳腺癌前哨淋巴结转移的成果，但准确率不高且对微转移情况检测效果不佳。针对以上问题，基于乳腺癌前哨淋巴结病理图像数据集（PCam），设计提出了SENet多路卷积神经网络模型。模型使用堆叠多路卷积单元和SENet模块，采用跳跃跨层连接、标准卷积与深度可分离卷积融合、加和与串联操作组合等策略。使用50%的图像迭代训练35次获得模型权重，然后采用准确率与AUC值指标对测试图像进行测试，准确率为97.32%，AUC值为98.05%。对比已有研究成果和主流卷积网络模型，该模型在49%、51%、100%测试集情况下，AUC值均排名第一。结果表明，该模型对淋巴结转移检测准确率较高，且对微转移也有很好的检测性能。

关键词: 乳腺癌, 前哨淋巴结转移, PCam, 多路卷积网络, SENet