基于Attention U-Net的乳腺X线图像微钙化检测模型的临床应用

Clinical application of mammogram microcalcification detection model based on Attention U-Net

目的:通过开发基于Attention U-Net的乳腺X线图像微钙化检测模型,实现微钙化的高效率检出,并探究不同性质钙化、不同乳腺密度对该深度学习模型微钙化检测性能的影响。方法:回顾性分析接受乳腺常规X线检查的347例患者的694幅图像。通过低年资医师独立阅片,高年资医师审核的方式,建立微钙化检出的参考标准。进行神经网络训练,建立深度学习模型。以钙化面积和数量分别计算,并采用精确率、召回率、F1分数、交并比等指标评估微钙化检测性能,分析不同性质钙化(良性vs恶性)、不同乳腺密度(a+b类vs c+d类)对深度学习模型微钙化检测性能的影响。结果:深度学习模型对微钙化检测的精确率为85.12%±18.39%(以钙化面积计算)和76.72%±19.85%(以钙化数量计算);召回率为78.18%±19.25%(以钙化面积计算)和85.12%±18.39%(以钙化数量计算);交并比为68.29%±21.39%(以钙化面积计算)和67.13%±23.84%(以钙化数量计算);F1分数为78.96%±17.70%(以钙化面积计算)和77.65%±9.37%(以钙化数量计算)。深度学习模型在不同钙化性质(良性vs恶性)中的精确率、召回率、交并比、F1分数之间差异均无统计学意义(P>0.05),在不同乳腺密度(a+b类vs c+d类)中对微钙化检测的精确率、召回率、交并比、F1分数之间差异无统计学意义(P>0.05)。结论:基于Attention U-Net的乳腺X线图像微钙化检测模型能够对乳腺微钙化进行有效的检测、有助于乳腺微钙化的定量研究,同时该模型稳定性强,钙化性质及乳腺密度对该模型的检测性能无影响。

Objective To develop a mammogram microcalcification detection model(DL model)based on Attention U-Net for realizing the efficient detection of microcalcifications,and to investigate the effects of breast density and microcalcification type on the microcalcification detection performance of the DLmodel.MethodsAretrospective analysis was performed on 694 images from 347 patients undergoing mammography.Through the independent image diagnosis by junior physicians and review by senior physicians,the reference standard for microcalcification detection was established.Neural network training was performed to establish a DL model.The performance of the model for microcalcification detection was evaluated using precision rate,recall rate,intersection over union(IoU)and F1-score which were calculated based on calcification area or quantity;and the effects of microcalcification type(benign vs malignant)and breast density(a+b vs c+d)on the model performance were also analyzed.Results For detecting microcalcifications by the DLmodel,the precision rate,recall rate,IoU and F1-score were 85.12%±18.39%,78.18%±19.25%,68.29%±21.39%and 78.96%±17.70%when the calculation was based on calcification area,and those were 76.72%±19.85%,85.12%±18.39%,67.13%±23.84%and 77.65%±9.37%when the calculation was based on calcification quantity.The differences in precision rate,recall rate,IoU,F1-score of DL model in different microcalcification types(benign vs malignant)and breast densities(a+b vs c+d)were insignificant.Conclusion The developed mammogram microcalcification detection model based on Attention U-Net can effectively detect breast microcalcifications and is conducive to the quantitative research on breast microcalcifications.Meanwhile,the model exhibits high stability,and the breast density and microcalcification type have trivial effects on the microcalcification detection performance of the model.