Artificial intelligence in pediatric osteopenia diagnosis: evaluating deep network classification and model...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Bone health research

Background:

• Osteopenia, a condition of low bone density, affects millions globally.
• Diagnosis typically relies on bone mineral density (BMD) assessments.
• Machine learning (ML) shows promise for medical image analysis.

Purpose of the Study:

• To evaluate deep learning networks for osteopenia classification using pediatric wrist X-rays.
• To apply explainable AI (XAI) for interpreting model decisions.
• To assess the potential of ML in clinical osteopenia diagnosis.

Main Methods:

• Utilized six deep learning networks (including CNNs and transformers) for binary classification (osteopenia vs. healthy).
• Employed the GRAZPEDWRI-DX pediatric wrist X-ray dataset.
• Applied two XAI techniques for visual explanation of model predictions.

Main Results:

• Deep networks effectively learned features distinguishing osteopenic from healthy bones.
• High classification accuracy rates were achieved across models.
• DenseNet201 with transfer learning achieved the highest accuracy at 95.2%.
• XAI provided interpretable insights into the models' decision-making processes.

Conclusions:

• Deep learning models demonstrate significant capability in differentiating osteopenia from healthy bone in pediatric wrist X-rays.
• The combination of high accuracy and interpretable explanations supports the integration of ML into clinical workflows.
• This approach holds promise for earlier and more accurate osteopenia diagnosis.