Leveraging Large-Scale Public Data for Artificial Intelligence-Driven Chest X-Ray Analysis and Diagnosis.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Machine Learning for Healthcare

Background:

• Chest X-ray (CXR) interpretation is vital for diagnosing thoracic diseases.
• Increasing demand for CXR analysis strains radiologist resources, especially in underserved areas.
• Developing automated diagnostic tools is crucial for efficient healthcare.

Purpose of the Study:

• To train general-purpose deep learning models for multi-label thoracic condition classification from CXRs.
• To evaluate the impact of data scale, diversity, and model architecture on diagnostic performance.
• To incorporate uncertainty quantification for assessing model reliability in clinical settings.

Main Methods:

• Trained multiple deep learning models (ResNet, DenseNet, EfficientNet, DLAD-10) on large, diverse public CXR datasets with noisy labels.
• Employed uncertainty quantification to gauge the reliability of model predictions.
• Validated model performance on internal and external datasets, analyzing effects of data scale and diversity.

Main Results:

• EfficientNet demonstrated superior performance, achieving the highest Area Under the ROC Curve (0.8944).
• Increased training data volume and diversity significantly improved diagnostic accuracy and generalizability.
• While larger datasets reduced predictive uncertainty, certain conditions like tuberculosis remained challenging due to data limitations.

Conclusions:

• General-purpose deep learning models can achieve reliable CXR diagnostic performance using large, diverse datasets, even with noisy labels.
• Data scale and diversity are key drivers for enhancing model generalizability and accuracy.
• Targeted strategies are necessary to address diagnostic challenges for underrepresented thoracic conditions.