摘要
目的:为解决传统机器学习在病理图像诊断方面的性能不足与纯粹人工阅片导致的错诊,以及传统的手动图形数据增强扩充对模型性能提高的未知性等问题,设计一个用于计算机辅助诊断(Computer-Aided Diagnosis, CAD)的乳腺癌病理图像自动分类模型。方法:结合深度学习在图像识别的优势,以残差网络为基础网络框架,使用迁移学习技术方法加快模型的收敛和训练,并使用自动增强(AutoAugment,AA)手段替代传统的数据增强手段,以实现提高模型性能的目的。结果:使用AA手段增强扩充的数据集训练出的模型,相对于未进行数据扩充以及进行传统手动扩充数据训练的模型其准确率均提升1个百分点,同时在测试集中的恶性肿瘤样本中,更是达到98.7%的灵敏度。结论:使用AA手段能有效提高模型的性能,为数据的扩充提供了新的技术方法,为提高模型识别性能提供了新的技术手段,同时也为CAD应用于实际临床诊断做了可行性论证。
Aims: A breast cancer histopathological images auto-classification model for computer-aided diagnosis (CAD) was designed to solve the problem of poor performance of traditional machine learning in pathological image diagnosis, misdiagnosis caused by pure manual reading and the unknown performance improvement of traditional manual graphic data enhancement. Methods: Combined with the advantages of deep learning in image recognition, the transfer learning technology method was used to accelerate the convergence and training of the model in which Residual network was used as the basic network framework. AutoAugment (AA) was used to replace traditional data enhancement methods to improve the performance of the model. Results: The accuracy of the model in which the AA method was used to extend data set rose 1 percentage point, compared to those without data extension and with traditional manual extension data. Besides, the sensitivity was up to 98.7% in the malignant tumor samples in the test set. Conclusions: The performance of the model can be effectively improved by using the AA method and a new technical method for data extension is provided. The model also provides a new traditional means for improving identification performance and is feasible for CAD application in practical clinical medical diagnosis.
作者
王恒
李霞
沈茜
徐文龙
WANG Heng;LI Xia;SHEN Xi;XU Wenlong(College of Information Engineering,China Jiliang University,Hangzhou 310018,China)
出处
《中国计量大学学报》
2019年第3期343-350,共8页
Journal of China University of Metrology
基金
国家自然科学基金项目(No.61672476)
关键词
计量
乳腺癌病理图像
残差网络
深度学习
自动增强
metrology
breast cancer histopathological image
residual network
deep learning
AutoAugment
