基于BIRADs多任务学习模型的乳腺肿块分类  

Breast mass classification based on BIRADs multi-task learning model

作　　者：吴书裕 周露 王琳婧 李慧君 张书旭 梅颖洁 WU Shuyu;ZHOU Lu;WANG Linjing;LI Huijun;ZHANG Shuxu;MEI Yingjie(Department of Radiotherapy,Affiliated Cancer Hospital and Institute of Guangzhou Medical University,Guangzhou 510095,China;Department of Radiology,Guangdong Provincial People's Hospital,Guangzhou 510080,China)

机构地区：[1]广州医科大学附属肿瘤医院放疗科,广东广州510095 [2]广东省人民医院放射科,广东广州510080

出　　处：《中国医学物理学杂志》2023年第10期1220-1227,共8页Chinese Journal of Medical Physics

基　　金：广东省自然科学基金(2020A1515110577);广东省医学科研基金(A2023291)。

摘　　要：目的:为解决乳腺图像肿块分类与深度学习应用的难题,提出一种基于乳腺影像报告与数据系统(BIRADs)多任务学习模型的肿块分类方法。方法:构建迁移学习的形态学和纹理特征提取器,并在此基础上引入多任务分类器,实现BIRADs诊断相关的边缘、形状、密度和微小性评估。研究通过训练策略、输入图像和模型架构系列实验和指标,分析评估模型性能。结果:在迁移学习策略下,Base模型和BIRADs模型性能均有显著提升。原始肿块图像作为输入的模型性能均优于掩模图像模型。在迁移学习和原始肿块输入下,BIRADs模型相较Base模型有更高的AUC值(0.830 vs 0.793)、准确率(0.747±0.024 vs 0.712±0.023)、精确率(0.643±0.032 vs 0.607±0.030)、召回率(0.774±0.037 vs 0.715±0.042)、F1-score(0.702±0.028 vs 0.656±0.029)。多任务学习模型在乳腺肿块分类中具有显著优势。结论:BIRADs多任务学习模型结合临床知识与数据驱动方法显著提高肿块分类准确性和模型鲁棒性,有望提高乳腺癌诊断准确性。Objective To propose a breast mass classification method based on multi-task learning model of breast imaging reporting and data system(BIRADs)for addressing the challenges in breast mass classification and deep learning applications.Methods A BIRADs multi-task learning model with a transfer learning-based morphological feature extractor,texture feature extractor,and multi-task classifier,which enables the assessment of BIRADs-related characteristics such as margin,shape,density and subtlety,was constructed.Model performance was analyzed through experiments involving training strategies,network architectures,and input types.Results With transfer learning as the training strategy,both Base and BIRADs models exhibit significant performance improvement.The models with original mass images as input outperformed those with masked images.With transfer learning as the training strategy and original mass images as input,the BIRADs model had higher AUC,accuracy,precision,recall rate,and F1-score as compared with Base model(0.830 vs 0.793,0.747±0.024 vs 0.712±0.023,0.643±0.032 vs 0.607±0.030,0.774±0.037 vs 0.715±0.042,0.702±0.028 vs 0.656±0.029,respectively).The multi-task learning model demonstrated significant advantages in breast mass classification.Conclusion The proposed BIRADs multi-task learning model combining clinical knowledge and data-driven methods enhances accuracy and robustness in breast mass classification,which is expected to improve the diagnostic accuracy of breast cancer.

关 键 词：乳腺癌 肿块分类 数据驱动 迁移学习 多任务学习 

分 类 号：R318[医药卫生—生物医学工程] R445[医药卫生—基础医学]