Leveraging Masked Autoencoders for Enhanced Classification of Multi-labeled Pediatric Thoracic Diseases in Chest X-rays

Pediatric thoracic diseases present significant health risks and are commonly diagnosed using chest X-rays. However, interpreting pediatric images faces numerous challenges. While deep learning shows promise as a solution, it suffers from the limited availability of pediatric X-ray data. To address these issues, we propose leveraging masked autoencoders (MAE) in a two-step training process: initially pretraining the MAE model on adult chest X-rays, then fine-tuning it with pediatric chest X-rays for the diagnosis of multi-labeled thoracic diseases. Our approach yielded superior performance metrics (AUC: 0.735, F1-score: 0.602), surpassing both ResNet-34 (AUC: 0.669, F1-score: 0.560) and ViT-S (AUC: 0.645, F1-score: 0.539) trained from scratch. Additionally, the MAE model outperformed ResNet-34 (AUC: 0.697, F1-score: 0.591) and ViT-S (AUC: 0.638, F1-score: 0.559), pretrained on the ImageNet dataset. Despite using substantially less X-ray data compared to the ImageNet dataset, the MAE-based approach demonstrated superior performance. These findings emphasize the effectiveness of our approach in addressing the complexities of pediatric thoracic disease diagnosis.