SAM-X: sorting algorithm for musculoskeletal x-ray radiography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Musculoskeletal image datasets are vast and unstructured, hindering deep learning applications.
• Efficiently organizing these datasets by anatomical entity is crucial for research and clinical use.

Purpose of the Study:

• To develop and validate a two-phased deep learning algorithm for sorting post-X-ray musculoskeletal images.
• The algorithm aims to classify radiographs by anatomical entity to facilitate large dataset construction.

Main Methods:

• A self-supervised model initially clustered 42,608 radiographs into 1000 groups.
• A radiologist assigned semantic labels, and a modified model was trained as a classifier.
• Validation included data splitting, cross-validation, and a 50% external hold-out test set. Grad-CAMs were used for interpretability.

Main Results:

• The self-supervised clustering achieved a normalized mutual information of 0.930.
• The final classifier reached 96.6% accuracy on the hold-out test set for top-class prediction.
• High accuracy (99.6%) was achieved when considering the top two predicted labels, demonstrating robustness.

Conclusions:

• The proposed deep learning algorithm accurately classifies radiographs by anatomical entity.
• This method is effective for building large, structured radiograph datasets for musculoskeletal disease assessment.
• The algorithm optimizes radiological workflows, increasing efficiency and reducing manual tasks.