摘要
目的 构建适应中国人群的乳腺钼靶图像分类辅助系统,探讨人工智能技术在国内辅助乳腺癌早期筛查的潜力.方法 为复现当前文献中的主流深度学习方法,利用乳腺X线筛查数字数据库子集(curated breast imaging subset of digital database for screening mammography,CBIS-DDSM)、乳房X线图像分析学会数据库(mammographic image analysis society database,MIAS)等国际公开数据集分别进行模型训练,并在华教科技有限公司提供的中国人群乳腺钼靶图像数据集(Chinese breast mammography dataset,CBMD)上进行测试和模型性能比较;针对中国人群数据在公开数据集训练模型性能测试不理想的问题,结合中国人群数据特点,提出基于滑动窗口调窗机制的优化策略,设计二阶段迁移学习方法,以提升模型的整体性能,并进行系统研发.结果 使用滑动窗口调窗机制及二阶段迁移学习后的CBMD训练模型,以中国人群数据集为测试集,其准确度从默认窗口下公开数据集训练模型的0.50提升至0.80,精度从0.54提升至0.82,灵敏度从0.52提升至0.80,F1值从0.52提升至0.80,AUC值从0.51提升至0.89.结论 本研究引入滑动窗口调窗机制和二阶段迁移学习策略,显著提升了乳腺钼靶图像分类模型在中国人群数据集上的性能,初步达到适应中国人群的乳腺钼靶图像辅助分类的目的.
Abstract
Objective To construct a mammography image classification assistant system suitable for Chinese population,and explore the potential of artificial intelligence technology to assist early screening of breast cancer in China.Methods Curated breast imaging subset of digital database for screening mammography(CBIS-DDSM),Mammographic image analysis society database(MIAS)and other international open datasets were used to conduct model training respectively in order to reproduce the mainstream in-depth learning methods in the current literature.The model was also tested on the Chinese breast mammography database(CBMD)provided by Huajiao Technology Co.,Ltd,and the performance was compared.Aiming at the problem that the Chinese population data are not ideal in the performance test of the open dataset training model,an optimization strategy based on the sliding window adjustment mechanism was implemented in combination with the characteristics of Chinese population data.Then a two-stage migration learning method was designed to improve the overall performance of the model,and then development of our system was carried out.Results With the sliding window adjustment mechanism and the CBMD training model after two-stage transfer learning,the accuracy of our developed system was improved from 0.50 of the open datasets to 0.80,precision from 0.54 to 0.82,sensitivity from 0.52 to 0.80,F1 value from 0.52 to 0.80,and AUC value from 0.51 to 0.89 based on the Chinese population dataset as the test set.Conclusion Through the introduction of sliding window adjustment mechanism and two-stage migration learning strategy,the performance of the breast molybdenum target image classification model has been significantly improved in the Chinese population dataset,and our system primarily achieves the purpose of assisting the classification of breast molybdenum target images for the Chinese population.
国家科技期刊平台
NSTL
维普
万方数据