A deep learning approach to automatically quantify lower extremity alignment in children.

Area of Science:

• Medical imaging and artificial intelligence
• Pediatric orthopedics
• Biomechanical analysis

Background:

• Accurate quantification of lower extremity alignment in children is crucial for diagnosing and managing orthopedic conditions.
• Traditional methods for calculating hip-knee-ankle angles (HKAAs) rely on manual landmark identification, which can be time-consuming and prone to variability.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for automated prediction of anatomical landmarks on pediatric lower extremity radiographs.
• To assess the CNN's ability to accurately calculate hip-knee-ankle angles (HKAAs) and quantify lower extremity alignment.

Main Methods:

• A CNN was trained on 528 full-length lower extremity radiographs from children (≤18 years old).
• Anatomical landmarks for HKAAs were manually identified by a radiologist and used as ground truth.
• The CNN predicted landmarks by regressing heatmaps, and the resultant HKAAs were compared to ground truth using absolute prediction error and intraclass correlation.

Main Results:

• The CNN demonstrated high accuracy in predicting HKAAs, with a mean absolute prediction error of 0.94° ± 0.84° after excluding outliers.
• An excellent intraclass correlation of 0.974 was observed between the CNN-predicted and ground truth HKAAs.
• Only 1.1% of predictions had an absolute error greater than 10°, indicating minimal outliers.

Conclusions:

• The developed CNN is a promising tool for accurately predicting anatomical landmarks and calculating HKAAs from pediatric radiographs.
• This AI-driven approach has the potential to serve as an effective computer-aided diagnostic tool for assessing pediatric lower extremity alignment.