Automated Analysis of Pelvic Radiographs for Hip Dysplasia Screening Using Artificial Intelligence in Children with Cerebral Palsy: A Systematic Review

TL;DR

This paper reviews AI models for detecting hip dysplasia in children with cerebral palsy, showing promising accuracy compared to experts.

Contribution

The study systematically evaluates AI performance in hip dysplasia detection, highlighting its potential as a clinical tool.

Findings

AI sensitivity for hip dysplasia detection ranged from 70% to 97.4%.

Specificity ranged from 85% to 96%, with high area under the curve values (0.923–0.999).

Most studies had moderate to high risk of bias due to internal validation and small datasets.

Abstract

Background and Objectives: Cerebral palsy is a debilitating and complex movement disorder affecting millions of people worldwide. Many children with cerebral palsy develop hip dysplasia, which can lead to pain, functional decline, and long-term complications. Regular hip surveillance is therefore essential to allow early intervention and prevent progression. At present, screening is performed manually by experienced clinicians, which can be time consuming and costly. This study aimed to compare the performance of artificial intelligence models with expert clinicians in detecting hip dysplasia in children with cerebral palsy. Materials and Methods: A thorough search of Embase, Ovid MEDLINE, and Web of Science was conducted from inception to July 2025. Studies evaluating AI-based detection of hip dysplasia in children aged 18 years or younger with cerebral palsy were included. Risk of bias was assessed using the QUADAS-2 tool. Results were synthesised narratively in accordance with SWiM guidelines. Results: Across the six included studies, which included over 4000 radiographs, AI sensitivity for detecting hip dysplasia ranged from 70% to 97.4%, and specificity ranged from 85% to 96%, depending on the migration percentage thresholds applied. Area under the curve values ranged from 0.923 to 0.999. Only one study performed external validation using a national surveillance dataset. Risk of bias was moderate to high in most studies due to internal validation and small datasets. Conclusions: The findings suggest that AI demonstrates potential as an adjunct for hip surveillance in children with cerebral palsy.

Figures