一种乳腺X线影像肿块征象检测算法

Algorithm for Detection of Mass Signs in Mammography

针对乳腺X线影像数据样本少、肿块类别间差异不明显、乳腺肿块背景复杂与组织对比度低,导致检测的精度低等问题,文章提出了一种乳腺X线影像肿块征象检测算法.首先在主干网络中的卷积层中添加高效注意力机制,提升算法对特征的提取能力;其次,在特征提取网络中引入非对称卷积结构,利用3种不同尺度的卷积核进行特征检测,增强模型对旋转和翻转目标的特征提取能力,最后使用One Cycle学习率调整策略在模型训练时跳出局部最优解.在DDSM(CBIS-DDSM)公开数据集测试,本文提出的YOLOv5-EA模型有效提高了乳腺X线影像目标检测精度,在识别X线影像的良、恶性肿块及钙化灶的准确率分别达到了93.0%,88.4%和88.1%.

This paper presents an algorithm for detecting the signs of breast masses in X-ray mammography,which aims at the problems of few samples of mammography data,no obvious difference between the types of tumors,complex background of mammography masses and low tissue contrast,resulting in low accuracy of detection.First,an efficient attention mechanism is added to the convolution layer in the main network to improve the algorithm's ability to extract features.Secondly,the asymmetric convolution structure is introduced in the feature extraction network,and the convolution kernels of three different scales are used for feature detection to enhance the feature extraction ability of the model for rotating and flipping targets.Finally,the One Cycle learning rate adjustment strategy is used to jump out of the local optimal solution during model training.In the open dataset test of DDSM(CBIS-DDSM),the YOLOv5-EA model proposed in this paper effectively improves the target detection accuracy of mammography.The accuracy of identifying benign and malignant tumors and calcified lesions on mammography is 93.0%,88.4%and 88.1%,respectively.