Enhancing pediatric pneumonia diagnosis through masked autoencoders.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Pulmonology

Background:

• Pneumonia diagnosis in children under 5 is critical due to immature immune systems.
• Chest X-rays are vital but challenging for pediatric pneumonia diagnosis due to subtle findings and interpretation subjectivity.
• Deep learning shows promise but is hindered by limited labeled pediatric chest X-ray data.

Purpose of the Study:

• To explore self-supervised learning, specifically Masked Autoencoder (MAE), to address data scarcity in pediatric pneumonia diagnosis.
• To enhance diagnostic accuracy for pediatric pneumonia by pretraining MAE on adult data and fine-tuning on pediatric data.
• To investigate the impact of masking ratios and labeled data efficiency of MAE in pediatric pneumonia detection.

Main Methods:

• Utilized a self-supervised learning approach with Masked Autoencoder (MAE).
• Pretrained the MAE model on adult chest X-ray images.
• Fine-tuned the pretrained MAE model on a pediatric pneumonia chest X-ray dataset for classification tasks.

Main Results:

• Achieved high performance in distinguishing normal from pneumonia cases (AUC: 0.996, Accuracy: 95.89%).
• Demonstrated strong classification accuracy for normal, bacterial, and viral pneumonia (AUCs: 0.997, 0.983, 0.956; Accuracy: 93.86%).
• Investigated the influence of masking ratios and evaluated the labeled data efficiency of the MAE model.

Conclusions:

• Self-supervised learning with MAE offers a promising solution to overcome data scarcity in pediatric pneumonia diagnosis.
• The proposed approach significantly enhances diagnostic capabilities for pediatric pneumonia using chest X-rays.
• MAE demonstrates potential for improving the accuracy and efficiency of AI-driven diagnostic tools in pediatric radiology.