Real-Time Ultrasound Diagnosis of Developmental Dysplasia of the Hip Using an Attention-Enhanced YOLOv11 Model

TL;DR

This paper introduces a fast and accurate AI model for real-time ultrasound diagnosis of hip dysplasia in children, reducing reliance on operator skill.

Contribution

An attention-enhanced YOLOv11 model for automated DDH classification with high accuracy and clinical interpretability.

Findings

The model achieved 95.05% accuracy with 11.5 ms inference speed, outperforming MobileNetV3 and ShuffleNetV2.

Grad-CAM visualizations showed the model focuses on clinically relevant landmarks like the acetabular roof and femoral head.

The framework combines technical robustness with clinical relevance for potential real-time use.

Abstract

Developmental dysplasia of the hip (DDH) is a common pediatric orthopedic disorder that can lead to lifelong disability if undetected. Ultrasound is the primary diagnostic modality but is subject to operator dependence and inter-observer variability. To address this challenge, we propose an attention-enhanced YOLOv11 framework for automated DDH classification. A dataset of 6,075 hip ultrasound images was preprocessed with augmentation and dimensionality reduction via UMAP. The model integrates Cross-Stage Partial (CSP) modules and C2PSA spatial attention to improve feature extraction, and was trained using Focal Loss and IoU Loss. It achieved 95.05% accuracy with an inference speed of 11.5 ms per image, substantially outperforming MobileNetV3 and ShuffleNetV2. Grad-CAM visualizations confirmed that the model consistently attends to the acetabular roof and femoral head, landmarks central to Graf classification, thereby enhancing clinical interpretability. These findings demonstrate that the proposed framework combines technical robustness with clinical relevance. Future work will emphasize multi-center validation and multimodal integration to ensure generalizability and support widespread clinical adoption.

Figures