Can images crowdsourced from the internet be used to train generalizable joint dislocation deep learning algorithms?

Area of Science:

• Orthopedic imaging analysis
• Artificial intelligence in medicine
• Deep learning for medical diagnosis

Background:

• Deep learning (DL) models can potentially automate the triage of orthopedic emergencies like joint dislocations.
• The rarity of these injuries often makes it difficult to gather sufficient data for training DL algorithms.
• This study investigates the feasibility of using internet-sourced images for training DL models for joint dislocation detection.

Purpose of the Study:

• To evaluate the efficacy of using internet-sourced radiographs to train Convolutional Neural Networks (CNNs) for identifying joint dislocations.
• To assess if these trained CNNs can generalize to real-world clinical scenarios.
• To determine if online repositories are a viable data source for rare orthopedic conditions.

Main Methods:

• Collected 100 radiographs (50 dislocated, 50 located) per joint (shoulder, elbow, hip, THA) from online repositories.
• Trained various CNN binary classifiers using on-the-fly and static data augmentation.
• Evaluated the best-performing CNNs on an external test set from three hospitals, using Area Under the Receiver Operating Characteristic Curve (AUROC) and Class Activation Maps (CAMs).

Main Results:

• The best CNNs achieved high AUROCs on both internal and external test sets for all evaluated joints (elbow, hip, shoulder, THA).
• External test set AUROCs ranged from 0.880 (hip) to 0.998 (elbow).
• CAM heatmaps confirmed that the CNNs focused on relevant anatomical areas for diagnosis.

Conclusions:

• Internet-sourced radiographs, even in modest quantities, can effectively train CNNs for diagnosing joint dislocations.
• These trained models demonstrate clinical generalizability.
• Online repositories offer a practical solution for acquiring training data for rare orthopedic conditions, facilitating AI development.