Simplified Transfer Learning for Chest Radiography Models Using Less Data.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Developing deep learning models for radiology demands extensive data and computational resources.
• Data limitations are worsened by shifts in patient populations and care standards, as seen during the COVID-19 pandemic.
• Traditional transfer learning pretrains models on nonmedical data, which is a partial solution.

Purpose of the Study:

• To decrease data set size requirements for deep learning models in chest radiography.
• To achieve this by employing supervised contrastive (SupCon) learning to generate robust chest radiography networks.

Main Methods:

• Generated chest radiography networks using SupCon learning on over 821,000 images from India and the US.
• Utilized these networks for 10 prediction tasks (e.g., tuberculosis, COVID-19 outcomes) on datasets from India, US, and China.
• Evaluated three model development setups (linear, nonlinear classifiers, full network fine-tuning) across various data set sizes.

Main Results:

• SupCon learning reduced label requirements up to 688-fold compared to nonmedical transfer learning.
• Achieved superior Area Under the Curve (AUC) performance at comparable data set sizes.
• In low-data scenarios (45 images), models achieved an AUC of 0.95 for tuberculosis classification, matching radiologist performance.

Conclusions:

• Supervised contrastive learning enables high-performance deep learning models with minimal data (as few as 45 images).
• This method is promising for predictive modeling with small data sets and adapting to evolving patient populations.
• SupCon learning offers a powerful alternative for developing AI in radiology, especially in data-scarce environments.