# Infant hip screening using multi-class ultrasound scan segmentation

**Authors:** Andrew Stamper, Abhinav Singh, James McCouat, Irina Voiculescu  
**arXiv:** [2211.04350](https://arxiv.org/abs/2211.04350)  
**Date:** 2022-11-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.04350v1/

## Key points

- Developmental dysplasia of the hip (DDH) affects infants where the femoral head is incorrectly located in the hip joint.
- Current state-of-the-art methods for DDH evaluation use the Graf method, which has high inter-operator variability and only achieves approximately 85% agreement with clinicians.
- A deep learning algorithm is proposed to segment key structures within ultrasound images to calculate Femoral Head Coverage (FHC) and provide a screening diagnosis for DDH.
- The algorithm employs a modified multi-class U-Net CNN to segment three key anatomical structures: ilium, femoral head, and labrum from static 2D ultrasound images.
- The model was trained on a small dataset of 200 scans using data augmentation techniques to reduce overfitting risks.
- The model outperformed the international state-of-the-art method, agreeing with expert clinicians on 89.8% of our test images.
- The proposed algorithm takes automated numerical measures of FHC (AFHC) to classify the presence or absence of DDH by evaluating FHC%, which is calculated as A/B x 100 (where A represents the area covered by femoral head and B represents the total area covered by acetabulum).
- This reproducible algorithm provides an accurate and reliable screening diagnosis for DDH, enabling early detection and intervention.
- This approach has significant potential in improving early detection rates and reducing inter-operator variability associated with current state-of-the-art methods.

## Summary

Developmental dysplasia of the hip (DDH) is a condition that affects infants where the femoral head is incorrectly located in the hip joint. Current state-of-the-art methods for DDH evaluation use the Graf method, which has high inter-operator variability and only achieves approximately 85% agreement with clinicians. To address this issue, we propose a deep learning algorithm for segmenting key structures within ultrasound images to calculate Femoral Head Coverage (FHC) and provide a screening diagnosis for DDH. This is the first study to automate FHC calculation for DDH screening. Our algorithm employs a modified multi-class U-Net CNN to segment three key anatomical structures: ilium, femoral head, and labrum from static 2D ultrasound images. The model was trained on a small dataset of 200 scans using data augmentation techniques to reduce overfitting risks. The model outperformed the international state-of-the-art method, agreeing with expert clinicians on 89.8% of our test images. The proposed algorithm takes automated numerical measures of FHC (AFHC) to classify the presence or absence of DDH by evaluating FHC%, which is calculated as A/B x 100 (where A represents the area covered by femoral head and B represents the total area covered by acetabulum). Our reproducible algorithm provides an accurate and reliable screening diagnosis for DDH, enabling early detection and intervention. In conclusion, our study presents an innovative approach to automate FHC calculation for DDH screening using deep learning algorithms that can accurately segment key anatomical structures in ultrasound images. This approach has significant potential in improving early detection rates and reducing inter-operator variability associated with current state-of-the-art methods.