基于多光谱通道注意力和双向特征融合的乳腺肿块检测
Breast mass detection based on multi-spectral channel attention and bidirectional feature fusion
投稿时间:2022-06-14  修订日期:2022-06-14
DOI:
中文关键词: 乳腺钼靶  肿块检测  RetinaNet  注意力机制  双向特征融合
英文关键词: mammography  mass detection  RetinaNet  attention mechanism  bidirectional feature fusion
基金项目:中央高校基本科研业务费专项资金项目(CZY22015)
作者单位邮编
陈罗林 中南民族大学 计算机科学学院 
周煜松 中南民族大学 计算机科学学院 
徐胜舟 中南民族大学 计算机科学学院 430074
摘要点击次数: 38
全文下载次数: 0
中文摘要:
      乳腺钼靶X线摄影是诊断乳腺疾病的有效手段,计算机辅助诊断系统在乳腺肿块检测中起着重要作用,针对检测过程中易出现漏检、误检导致精度不理想的情况,本文提出一种改进的RetinaNet乳腺肿块检测算法。首先,在RetinaNet的基础上对特征提取部分进行改进,在每个残差块中引入多光谱通道注意力机制,使网络能更加关注到目标区域;然后在特征融合部分,以两种不同方式添加一条自底向上的路径,并通过横向连接最终实现深浅层特征的双向融合,加快浅层信息流通,使得到的特征信息更加丰富。实验表明,改进后的算法在乳腺钼靶图像的肿块检测中具有良好的检测效果,既减少了漏检率,平均精度也提升了3.2%,并且相比其他出色的检测算法,也具有一定的优势。
英文摘要:
      Mammography is an effective method for diagnosing breast diseases, and the computer aided diagnostic system plays an important role in the detection of breast masses. Because it is prone to the situation that missing and false detections lead to suboptimal accuracy, this paper proposes an improved RetinaNet algorithm of breast mass detection. Firstly, on the basis of RetinaNet algorithm, the part of the feature extraction is improved. The multi-spectral channel attention mechanism is introduced into each residual block, so that the network can pay more attention to the target area. Then in the part of feature fusion, a bottom-up path is added in two different ways, and the bidirectional fusion of deep and shallow features is finally achieved through horizontal connections. It can speed up the flow of shallow information and make the obtained feature information more abundant. Experiments show that the improved algorithm has a good detection capability in the mass detection of mammography images. It not only reduces the missed detection rate, but also improves the average precision by 3.2%. Compared with other excellent detection algorithms, it also has certain advantages.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭