Poster 154. A Novel Machine Learning Algorithm for the Establishment of Pediatric Bone Age Using Knee Radiographs

Résumé fourni par la source

Objective: Bone age assessment is a critical tool in pediatric healthcare for evaluating skeletal maturation and aiding in the diagnosis and management of endocrine disorders, skeletal anomalies, and metabolic conditions. In orthopedic surgery, bone age estimation is used for perioperative surgical planning including decision to pursue surgery, the type of procedure offered, and the surgical technique employed. Anterior-posterior left hand radiographs have traditionally been used to estimate skeletal age for decades, however, this practice requires additional imaging beyond the clinical indication. Advances in computer vision techniques, particularly convolutional neural networks, can detect subtle skeletal patterns that may not be readily apparent to human observers. Leveraging these techniques, the aim of this study was to develop a deep learning model capable of estimating bone age directly from knee radiographs, reducing the need for hand films. Methods: 7,336 knee radiographs from 5,701 patients under 18 years of age and obtained between January 2018 and January 2024 were analyzed. The images included a range of normal images and images with pathology from multiple radiographic views including right two-view (1,167), left two-view (1,252), right three-view (768), left three-view (831), right four-plus view (1,282), and left four-plus view (1,280). Data were split into training (80%), tuning (10%), and testing (10%) sets. Patients with more than one study were placed in the same cohort in order to prevent data leakage. Results: We developed a view-agnostic multimodal deep learning model using an intermediate fusion approach. Our model employed a 2D DenseNet121 as the imaging feature extractor and two shallow neural networks (Figure 1). Multiple radiographic views of the knee are processed through a convolutional neural network, which functions as the image feature extractor and generates ten imaging features that capture relevant patterns of skeletal development. In parallel, patient sex is processed through a first shallow neural network. The second neural network merged and processed the imaging features along with the transformed sex features to allow for concatenation and a prediction of skeletal age in months. Conclusions: This study demonstrates that automated skeletal age estimation from knee radiographs is feasible using a multimodal deep learning model, with performance comparable to hand radiograph-based bone age methods. To our knowledge, this is the first study to leverage multiple knee radiographic views for bone age determination. The slightly higher mean absolute error observed may reflect greater variability in training data or a smaller dataset. Future study should include external validation of the model on outside datasets, prospective clinical application, and direct comparison of model predictions with radiologist atlas-based assessments to further evaluate precision.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.