Deep Learning-Based Automatic Diagnosis System for Developmental Dysplasia of the Hip

Area of Science:

• Artificial Intelligence in Medicine
• Radiology
• Orthopedics

Background:

• Clinical diagnosis of developmental dysplasia of the hip (DDH) relies on manual measurement of radiological angles (Center-Edge, Tönnis, Sharp) from pelvic radiographs.
• Manual measurements are time-consuming and prone to inter-observer variability, impacting diagnostic accuracy and consistency.

Purpose of the Study:

• To develop and validate an automated deep learning system for accurate and consistent measurement of key radiological angles in DDH diagnosis.
• To create a novel data-driven scoring system for an explainable and comprehensive diagnostic output for DDH.

Main Methods:

• An end-to-end deep learning model was developed for keypoint detection on pelvic radiographs to automatically calculate CE, Tönnis, and Sharp angles.
• A data-driven scoring system was integrated to combine angle measurements into a single diagnostic output.
• The system's performance was evaluated against a cohort of orthopedists and individual angle criteria.

Main Results:

• The automated system demonstrated high consistency in angle measurements, with intraclass correlation coefficients of 0.957 (CE), 0.942 (Tönnis), and 0.966 (Sharp).
• The system achieved a diagnostic F1 score of 0.863, significantly outperforming the orthopedist group (0.777) and individual angle criteria.
• The AI system provided reliable, consistent, and explainable diagnostic outputs for DDH.

Conclusions:

• The developed AI system offers a reliable and consistent method for automated radiological angle measurements in DDH diagnosis.
• This AI-powered tool enhances diagnostic accuracy and reduces variability compared to manual methods, providing clinicians with a more interpretable solution.
• The system has significant clinical impact by improving the consistency and interpretability of DDH diagnosis.