一种双视图信息融合的乳腺肿块自动检测算法

An Automatic Breast Mass Detection Algorithm with Dual-view Information Fusion

针对基于单视图的深度学习乳腺肿块检测算法假阳性率较高问题,提出一种双视图信息融合的乳腺肿块自动检测算法。首先建立双曲正割模型,利用互相关法自动搜索乳腺X线摄影图像中的肿块感兴趣区域,并根据胸壁线、乳头位置在双侧头尾位和内外侧斜位图像上建立物理坐标系,筛选标注假阳性区域以在数据层扩充监督学习信息;其次,设计空间金字塔池化模块有效融合基于YOLOv3主干网络提取的多尺度局部特征以提高检测敏感性;最后,在类别损失函数中增加聚焦参数,通过调节算法学习过程以提高检测特异性。充分利用双视图数据提供的先验信息提高检测正确率,实验结果表明检测敏感性达到92.0%,特异性达到87.7%,平均每幅图像假阳性0.041个,其检测性能较原模型大幅提升,且具有较好的鲁棒性。

Aiming to reduce the high false positive rate of breast mass detection algorithm based on single view deep learning,an automatic breast mass detection algorithm with dual-view information fusion was proposed.Firstly,Sech model was established and the cross-correlation method was used to automatically search the region of interest in breast X-ray images.According to the chest wall line and nipple position,a physical coordinate system was established on bilateral craniocaudal and mediolateral oblique images to screen and label the false positive areas to expand the supervised learning information in the data layer.Furthermore,the spatial pyramid pooling(SPP)was designed to effectively fuse the multi-scale local features extracted from the you only look once V3(YOLOV3)backbone network to improve the detection sensitivity.Finally,focusing parameters were added to the category loss function to improve the specificity of detection by adjusting the algorithm learning process.The proposed method made full use of the prior information provided by the dual-view data to improve the detection accuracy,the method achieved a sensitivity value of 92.0%,and specificity value of 87.7%,with 0.041 FP/I.Compared with the original model,the detection performance of the proposed model was greatly improved and has good robustness.