基于改进U-Net的髋关节关键点检测算法

Improved Algorithm for Keypoints Detection of Hip Based on U-Net

使用骨盆X光片诊断发育性髋关节发育不良(Developmental Dysplasia of the Hip,DDH)要求准确地标注髋关节关键点,而深度学习方法能作为可靠的辅助工具。针对骨盆片拍摄姿势和拍摄距离多样化问题,本文基于U-Net提出了RKD-UNet来检测髋关节关键点。该模型使用残差块改进U-Net的卷积层和skip-connection路径,并将坐标注意力引入到编码器中以增强模型对关键点邻域的特征提取能力。在编码器顶部使用卷积和ASPP模块构成Bridge块,以[3,6,9]的空洞率融合不同尺度的特征信息并提升模型的感受野。本文使用包含骨盆正位片、蛙位片、下肢全长片和术后骨盆片的数据集训练和测试模型。RKD-UNet实现了3.19±2.19 px的平均关键点检测误差和2.83°±2.59°的平均髋臼角测量误差。对正常、轻度、中度和重度脱位案例诊断的F1分数分别达到89.6、77.1、57.9和94.1,高于医生的手动诊断结果。实验结果表明,RKD-UNet能准确检测髋关节关键点并辅助医生诊断DDH。

The diagnosis of developmental dysplasia of the hip(DDH)using pelvic X-ray requires accurate mapping of hip key points,and deep learning methods can be used as reliable auxiliary tools.In order to solve the problem of diversified shooting posture and shooting distance for pelvic radiographs,this paper proposed RKD-UNet based on U-Net to detect keypoints of the hip.The model used residual blocks to improve U-Net’s convolution layers and skip-connection paths,as well as introduced the coordinate attention module into the encoder to enhance feature extraction ability for the keypoints neighborhood.Convolution layers and ASPP module were used on top of the encoder to form a Bridge block to fuse feature information at different scales and enhance the receptive field of the model with an atrous rate of[3,6,9].The model was trained and tested using radiographic data containing types of pelvic orthostasis,frog,full-length lower extremity,and postoperative pelvis.RKD-UNet achieves an average keypoints detection error of 3.19±2.19 px and an average acetabular angle measurement error of 2.83°±2.59°.The F1 score for the normal,mild,moderate,and severe dislocation cases were 89.6,77.1,57.9,and 94.1,respectively,which were higher than the doctors’diagnostic results.Experiments have shown that RKD-UNet can accurately detect keypoints of the hip and assist doctors in diagnosing DDH.