Machine Learning for Treatment Decision in Pediatric Forearm Fractures

Overview

This study developed a self-supervised vision transformer model pretrained on large public X-ray datasets and fine-tuned on a pediatric forearm fracture dataset to classify treatment recommendations as reduction or no-reduction needed. The model leverages unlabeled data for pretraining and achieves automated treatment decision support, addressing a gap in AI applications beyond fracture detection.

Background

Pediatric forearm fractures are common in emergency settings, with millions of cases annually in the USA. Treatment varies from immobilization to reduction or surgery depending on fracture displacement and characteristics. Many children are initially seen in non-specialized centers, leading to frequent transfers to pediatric orthopedic specialists, which can cause delays, increased costs, and burdens. Existing AI models focus mainly on fracture detection rather than guiding treatment decisions. Self-supervised learning with vision transformers offers a promising approach to leverage large unlabeled datasets for improved image analysis in this context.

Data Highlights

Key Findings

• Developed a self-supervised learning (SSL) pretraining model using vision transformers on 47,116 public X-ray images.
• Fine-tuned the pretrained model on an in-house pediatric forearm fracture dataset of 1,250 patients to classify treatment needs.
• Classification categories were “no-reduction needed” and “reduction needed” to guide clinical decision-making.
• SSL approach using group masked model learning (GMML) enabled learning from unlabeled data by reconstructing corrupted images.
• Model addresses the gap in AI applications by moving beyond fracture detection to treatment recommendation.
• Use of transformer architecture with attention mechanisms improved image representation learning for pediatric fracture assessment.

Clinical Implications

This automated framework can assist clinicians in making timely and accurate treatment decisions for pediatric forearm fractures, potentially reducing unnecessary transfers and associated costs. By integrating AI-driven treatment recommendations, primary care and urgent care providers may better manage cases locally, improving patient outcomes and resource utilization.

Conclusion

The study demonstrates the feasibility of using self-supervised vision transformers to classify treatment recommendations for pediatric forearm fractures from X-rays. This approach represents a significant advancement in AI-assisted fracture management by providing actionable clinical guidance beyond fracture detection.

References

1. Practice Management Committee of POSNA and AAP -- Referral Management in Pediatric Orthopedics
2. MURA Dataset -- Large Public Musculoskeletal Radiograph Dataset
3. FracAtlas Dataset -- Annotated Fracture Dataset
4. UNIFESP Dataset -- Medical Imaging Dataset
5. Self-Supervised Learning and Vision Transformers -- Recent Advances in Medical Imaging AI