Automatic Segmentation of the Paediatric Femoral Head

Area of Science:

• Orthopedic imaging
• Pediatric radiology
• Medical image analysis

Background:

• Developmental dysplasia of the hip (DDH) affects 0.1-3.4% of infants.
• Current DDH assessment is subjective, necessitating objective metrics.
• Femoral head morphology analysis can aid in objective DDH evaluation.

Purpose of the Study:

• To segment the pediatric femoral head in stable hips from radiographs for objective DDH assessment.
• To develop and evaluate novel segmentation techniques for pediatric hip joints.
• To establish a reliable method for analyzing femoral head morphology in DDH.

Main Methods:

• Comparison of a baseline U-Net model with data augmentation and region-of-interest (ROI) networks.
• Development of four models: baseline, data augmentation only, ROI only, and both techniques.
• Evaluation using tenfold cross-validation on 720 hip radiograph images.

Main Results:

• The U-Net model incorporating both data augmentation and ROI techniques achieved the highest performance.
• Achieved a Dice Similarity Coefficient (DSC) of 0.951±0.037, indicating excellent segmentation accuracy.
• This combined approach demonstrated superior results compared to baseline and individual techniques.

Conclusions:

• The developed segmentation algorithm provides an objective method for assessing pediatric hip joint morphology.
• This technique has the potential to improve early diagnosis and treatment planning for developmental dysplasia of the hip.
• Future work will leverage this algorithm to quantify hip joint morphology and assess the impact of early surgical intervention in DDH.