应用DeeplabV3+网络实现小儿髋关节超声图像识别

利用Graf法进行发育性髋关节发育不良(Developmental Dysplasia of the Hip,DDH)诊断时主要依靠骨-软骨交界面、股骨头、滑膜皱襞、关节囊及软骨膜、盂唇、软骨顶、骨性顶这7个解剖结构进行解剖验证，而初级医生对上述结构识别困难，因此文章提出了一种基于DeeplabV3+的网络模型用于7个结构的分割识别。首先对纳入的106例图像进行手动标记和预处理，之后将其分别输入DeeplabV3+和U-Net两种网络模型中，最终对其预测图表现和分割性能进行比较。与目前DDH图像分割中常用且表现优越的U-Net网络相比，DeeplabV3+网络的预测图包含的结构较多，边界分割也较清晰，其图像分割评价指标如相似性系数、豪斯多夫距离和平均豪斯多夫距离平均值的表现也优于U-Net网络。文章利用DeeplabV3+网络实现了DDH超声图像的7个结构分割，对临床医生进行后续图像的角度测量和分型诊断具有重要意义。

Application of DeeplabV3+ network in ultrasonic image recognition of pediatric hip joint

The diagnosis of developmental dysplasia of the hip (DDH) by Graf method mainly depends on the seven key structures, involving chondro-osseous border, femoral head, cartilagineous roof, synovial fold, labrum, bony roof and joint capsule. However, it is difficult for junior doctors to identify these structures. Therefore, a network model based on Deeplabv3+ for the segmentation of these seven structures is proposed in this paper. Firstly, 106 images were manually labeled and preprocessed, and then they were input into Deeplabv3+ and U-net network models respectively. Finally, the segmentation performances were compared. Compared with U-Net network which is commonly used and well-behaved in DDH segmentation, the image prediction of Deeplabv3+ network contain more structures and clearer boundary, and the main evaluation indices of segmentation, such as the average values of dice similarity coefficient, Hausdorff distance, average Hausdoff distance, also showed a better performance. The Deeplabv3+ network is first used to achieve segmentation of seven structures in DDH ultrasound images, which is of great significance for angle measurement and classification diagnosis.