AutoDDH: A dual-attention multi-task network for grading developmental dysplasia of the hip in ultrasound images

Mengyao Liu, Ruhan Liu, Jia Shu, Qirong Liu, Yuan Zhang, Lixin Jiang

Abstract: Developmental dysplasia of the hip (DDH) in newborns can cause serious long-term adverse effects if not promptly diagnosed and treated. Early intervention via ultrasound screening at 0–6 months is beneficial. However, early DDH diagnosis by ultrasound is complex and requires high-level experience for radiologists. Hence, deep learning for clinically assisted DDH screening is meaningful. The DDH classification task is challenging due to low ultrasound image resolution and difficulty in extracting structural features. We propose a dual-attention multi-task network (AutoDDH) for DDH grading using ultrasound images. It includes a dual-attention module for feature enhancement, a feature fusion module for detail improvement, and a dual-output branch for position embedding and generating outputs of DDH grading and anatomical structure segmentation. With the help of the segmentation task, the average accuracy and AUC of DDH four classifications reached 80.43\(\%\) and 0.96, outperforming other methods and laying the foundation for DDH intelligent assisted screening. Code available at: https://github.com/Liuruhan/AutoDDH.

External IDs:dblp:journals/vc/LiuLSLZJ25