Deploying deep learning models on unseen medical imaging using adversarial domain adaptation.

Area of Science:

• Machine Learning
• Medical Imaging
• Computer Vision

Background:

• Ensuring machine learning models generalize to new data is a key challenge.
• Dataset shift can degrade model performance on unseen data.
• Generative adversarial networks (GANs) offer a potential solution.

Purpose of the Study:

• To develop and evaluate a general technique using GANs to mitigate dataset shift.
• To improve the generalization of machine learning models on diverse datasets.
• To assess the efficacy of adversarial domain adaptation in medical imaging.

Main Methods:

• Applied generative adversarial networks (GANs) for domain adaptation.
• Utilized datasets of handwritten digits (149,298) and chest radiographs (868,549).
• Assessed performance by comparing Area Under the Curve (AUC) pre- and post-adaptation.

Main Results:

• Digit recognition: Baseline CNN AUC decreased on external data, but adaptation salvaged 35%.
• Radiograph classification: Baseline CNN AUC improved with adaptation, salvaging 25%.
• Adversarial domain adaptation significantly improved model performance on out-of-sample radiographic data.

Conclusions:

• Adversarial domain adaptation using GANs effectively ameliorates dataset shift.
• The technique enhances machine learning model performance on diverse medical imaging datasets.
• This approach can be integrated into machine learning deployment toolkits for medicine.