Improving Pediatric Pneumonia Diagnosis with Adult Chest X-ray Images Utilizing Contrastive Learning and Embedding...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Radiology

Background:

• Deep learning computer-aided diagnosis (CAD) for pneumonia in adult chest x-rays (CXRs) is advanced, but performance on pediatric CXRs is suboptimal due to limited annotated pediatric data.
• Leveraging large-scale adult CXR datasets is crucial for enhancing pediatric pneumonia detection capabilities.

Purpose of the Study:

• To develop a deep learning framework that effectively utilizes both adult and pediatric CXR datasets to improve pneumonia detection performance in pediatric populations.
• To address the domain shift challenge between adult and pediatric imaging data.

Main Methods:

• A three-branch parallel path learning framework was proposed, with paths trained on pediatric-only, adult-only, and combined CXRs.
• The framework employed multi-positive contrastive loss for classwise embedding clustering and embedding similarity loss to minimize domain shift effects.

Main Results:

• The proposed framework achieved a superior Area Under the Receiver Operating Characteristic (AUROC) score of 0.8464 on pediatric test datasets.
• This performance surpasses the conventional joint training approach, which yielded an AUROC of 0.8348.

Conclusions:

• The developed framework demonstrates superior performance in pediatric pneumonia detection by effectively integrating adult and pediatric CXR data.
• This approach offers a pathway towards generalized CAD models applicable to both adult and pediatric age groups, overcoming data scarcity issues in pediatric imaging.