CBA-YOLOv5s: A hip dysplasia detection algorithm based on YOLOv5s using angle consistency and bi-level routing attention

作者全名：Lv, Jia; Che, Junliang; Chen, Xin

作者地址：[Lv, Jia; Che, Junliang] Chongqing Normal Univ, Coll Comp & Informat Sci, Chongqing 401331, Peoples R China; [Lv, Jia] Chongqing Normal Univ, Natl Ctr Appl Math Chongqing, Chongqing 401331, Peoples R China; [Chen, Xin] Chongqing Med Univ, Childrens Hosp, Natl Clin Res Ctr Child Hlth & Disorders, Chongqing Key Lab Pediat,Minist Educ,Key Lab Child, Chongqing 400014, Peoples R China

通信作者：Chen, X (通讯作者)，Chongqing Med Univ, Childrens Hosp, Natl Clin Res Ctr Child Hlth & Disorders, Chongqing Key Lab Pediat,Minist Educ,Key Lab Child, Chongqing 400014, Peoples R China.

来源：BIOMEDICAL SIGNAL PROCESSING AND CONTROL

ESI学科分类：ENGINEERING

WOS号：WOS:001245997900001

JCR分区：Q1

影响因子：4.9

年份：2024

卷号：95

期号：　

开始页：　

结束页：　

文献类型：Article

关键词：Developmental Dysplasia of the Hip; Keypoint Detection; YOLOv5s; Angle Consistency; Bi-Level Routing Attention

摘要：The precise positioning of the hip joint keypoint is a prerequisite for accurately diagnosing hip dysplasia. However, there are still some disadvantages in existing methods, such as missed detection on hip joint keypoint, imprecise positioning, as well as insufficient constraints to the acetabular index (AI). Thus, it is a challenge to design an effective algorithm for avoiding missed detection and precisely locating the hip joint keypoint. In this paper, on the basis of You Only Look Once version 5 small (YOLOv5s), a novel hip dysplasia detection algorithm using angle consistency (AC) and bi-level routing attention (BRA) is proposed. Firstly, since the local neighborhood contains the local feature information of the hip joint keypoint, the detection of the hip joint keypoint is converted into its local neighborhood detection, and the influence of different local neighborhood sizes on the missed detection is explored to solve the missed detection problem. Secondly, a BRA module is introduced into the backbone network of YOLOv5s, so that the network can focus on the appropriate background pixels to reduce the impact of complex background on local neighborhood positioning. Finally, an AC loss function is designed to constrain the generation direction of local neighborhood. This algorithm is tested on a clinical dataset provided by the Department of Radiology, Children's Hospital of Chongqing Medical University. The experiment results show that the presented algorithm outperforms the state-of-the-art algorithms in many metrics. For example, the average error of L1 distance and the average angular error of acetabular index are 5.5124 pixels and 1.0481 degrees, reduced by 0.5626 pixels and 0.5361 degrees respectively, compared with YOLOv5s. Our finding indicates that it is an effective way to solve the missed detection by expanding the local neighborhood size. In addition, our exploration also suggested that filtering out irrelevant regions during attentional interaction contributes to the precise positioning of local neighborhoods. Moreover, the angular error of acetabular index can be reduced by designing the loss function to constrain the direction of the local neighborhood. Obviously, the algorithm can provide the possibility for automatic diagnosis of hip dysplasia.

基金机构：Chongqing Municipal Education Commission; Children's Hospital of Chongqing Medical University

基金资助正文：The authors declare the following financial interests/personal re- lationships which may be considered as potential competing interests: [Jia Lv reports financial support was provided by Chongqing Municipal Education Commission. Xin Chen reports financial support was provided by Children's Hospital of Chongqing Medical University.] .