Research on Cephalometric Landmark Points Detection Based on Dual Attention Mechanism
周金保 1武秀萍 2都冰丽 2张光华 3王烽飞 1卓广平 1马非3
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 太原师范学院计算机科学与技术学院,山西晋中 030619
2. 山西医科大学 口腔医院,山西太原 030001
3. 太原学院计算机科学与技术系,山西太原 030032
折叠
摘要
准确可靠的头影标志点检测在口腔正畸的临床诊疗和研究中至关重要,但在实践中存在标志点定位难度大、准度低以及速度慢等问题,因此为了降低检测难度,提高临床诊断的准确性和高效性,提出了一种结合多注意力机制的检测算法CenterNetSC.算法首先采用深度聚合网络DLA-34作为CenterNetSC的主干网络并用于特征提取;其次,通过在深度聚合网络中引入SE和CBAM两种注意力机制加强网络对卷积通道以及空间位置的全局感知;再次,在DLA-34网络之后加入可变形卷积捕捉对象的细节和局部信息;最后,模型在ISBI 2015 Grand Challenge的cephalometric X-rays两个测试集上分别实现了 1.11 mm和1.37 mm的平均径向误差(MRE),以及2.0 mm定位误差范围内87.13％和77.03％的成功检测率(SDR).较其他检测方法而言,Cen-terNetSC 能够快速、准确地定位标志点,可以满足临床医学的需求.
Abstract
Accurate and reliable detection of cephalometric landmark points is crucial in clini-cal diagnosis,treatment and research of orthodontics,but in practice there are problems such as difficulty in locating landmark points,low accuracy and slow speed.Therefore,in order to re-duce detection complexity and enhance the accuracy and efficiency of clinical diagnosis,a detec-tion algorithm called CenterNetSC,which incorporates multiple attention mechanisms,has been proposed.The algorithm first employs the Deep Layer Aggregation network(DLA-34)as the backbone network CenterNetSC and is used for feature extraction;secondly,two attention mech-anisms,SE and CBAM,are introduced into the Deep Layer Aggregation network to enhance the network's global perception of convolution channels and spatial positions;thirdly,adding de-formable convolutions after the DLA-34 network to capture the details and local information of the object.Finally,the model achieved mean radial errors(MRE)of 1.11 mm and 1.37 mm re-spectively on the two test sets of cephalometric X-rays of the ISBI 2015 Grand Challenge,and successful detection rates of 87.13％and 77.03％within an error range of 2.0 mm(SDR).Com-pared with other detection methods,CenterNetSC can quickly and accurately locate landmarks,which can meet the needs of clinical medicine.
维普
万方数据