基于多示例学习的钼靶图像BI-RADS分类方法

Multiinstance Networks for Mammogram BIRADS Classification

目的:基于深度学习的乳腺X线图像分类模型对于所得出的决策无法给出诊断依据。本文旨在保证模型分类精度的同时,改善模型的可解释性。方法:基于多示例学习,提出一种弱监督病灶分类和定位方法,以应用于乳腺X线图像的乳腺影像报告和数据系统(BI-RADS)分级任务。为了解决传统多示例学习只能应用于二元分类任务的问题,本文利用了BI-RADS分类的有序性,引入了新的样本标签、训练方法和输出判断方法。结果:公开数据集INbreast上的实验结果表明,在使用相同的特征提取网络下,所提出的方法在分类准确性上相较传统方法提升了2.91%。模型在病灶检出上的准确率达到83.49%,真阳性率达到75.87%,具有一定的病灶定位能力。结论:在不借助病灶轮廓数据或是位置数据进行训练的情况下,所提出的深度学习模型可以展示出每一个可能存在的病变区域及其BI-RADS分类,具有较好的应用场景。

Purpose:Mammogram classification model based on deep learning cannot be used to provide diagnostic basis for the obtained decisions.This paper aimed to meet the classification accuracy of the model and improve the interpretability of the model.Methods:A weakly supervised lesion classification and localization method based on multi-instance learning was proposed in this paper,and was applied to the Breast Imaging Reporting and Data System(BI-RADS)classification for mammography image.Based on the orderliness of BI-RADS levels,new sample labels,training methods,and output judgment methods were introduced to overcome the limitation of traditional multi-instance learning which could only be applied to binary classification.Results:The experimental results on the public dataset INbreast showed that using the same feature extraction network,the classification accuracy of the method in this paper was improved by 2.91%compared with the traditional method.The accuracy rate of the model in lesion detection reached 83.49%,and the true positive rate reached 75.87%.Conclusions:Without training on lesion segmentation data or location data,the model can display each possible lesion area and its BI-RADS level,which has a good application scenario.