Deep Mining External Imperfect Data for Chest X-Ray Disease Screening.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning excels in Chest X-ray (CXR) analysis but requires extensive, diverse data.
• Integrating multiple CXR datasets is crucial but challenging due to domain and label discrepancies.
• Existing methods struggle with joint training on heterogeneous CXR datasets.

Purpose of the Study:

• To develop a framework that effectively integrates external CXR data for improved thoracic disease classification.
• To address challenges of domain and label discrepancies in multi-dataset CXR analysis.
• To enhance the performance of deep learning models in classifying thoracic diseases.

Main Methods:

• Formulated multi-label thoracic disease classification as weighted independent binary tasks.
• Employed task-specific adversarial training to mitigate domain discrepancy (feature differences across datasets).
• Utilized uncertainty-aware temporal ensembling to handle label discrepancy (partially labeled data).

Main Results:

• The proposed framework successfully models and tackles both domain and label discrepancies.
• Achieved state-of-the-art performance on the NIH test set with an AUC of 0.8349.
• Demonstrated superior knowledge mining ability by effectively utilizing external CXR datasets.

Conclusions:

• The developed method significantly improves thoracic disease classification by leveraging multi-dataset CXR data.
• The framework offers a robust solution for challenges in medical image analysis with heterogeneous data.
• This approach advances the application of deep learning in automated medical diagnostics.