Enhancing thoracic disease detection using chest X-rays from PubMed Central Open Access.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Informatics

Background:

• Existing chest X-ray (CXR) datasets for training deep learning models are often limited by single-center origins and imbalanced pathology representation.
• Developing robust AI for thorax pathology detection requires diverse and comprehensive CXR data.

Purpose of the Study:

• To automatically construct a public, weakly-labeled CXR database from PubMed Central Open Access (PMC-OA) articles.
• To evaluate the impact of this novel database as supplementary training data for CXR pathology classification models.

Main Methods:

• A framework was developed for automated text extraction, CXR pathology verification, subfigure segmentation, and image modality classification from PMC-OA articles.
• The utility of the generated database was validated on thoracic diseases with historically poor performance in existing datasets (Hernia, Lung Lesion, Pneumonia, Pneumothorax).

Main Results:

• Classifiers fine-tuned with the automatically extracted PMC-CXR data showed significant performance improvements across all tested pathologies compared to models trained without it (e.g., Hernia AUC: 0.9335 vs 0.9154, p<0.0001).
• The framework successfully automates the collection of figures and associated legends, outperforming previous manual submission methods.
• Enhanced subfigure segmentation and advanced NLP techniques for pathology verification were incorporated.

Conclusions:

• The proposed automated framework effectively creates a valuable, large-scale, weakly-labeled CXR dataset from open-access literature.
• This approach significantly enhances the performance of deep learning models for CXR pathology detection, addressing data limitations in existing resources.
• The framework promotes findability, accessibility, interoperability, and reusability of biomedical image data.