Domain generalization on medical imaging classification using episodic training with task augmentation | JoVE Visualize

Area of Science:

Artificial Intelligence
Medical Imaging
Machine Learning

Background:

Medical imaging datasets often have domain shift due to variations in scanners and protocols, impacting machine learning model generalization.
Domain generalization (DG) aims to train models on diverse source domains for effective performance on unseen target domains, crucial for medical applications.
Existing meta-learning approaches like MAML struggle with limited source domains, risking overfitting in medical DG tasks.

Purpose of the Study:

To propose a novel domain generalization (DG) scheme for medical imaging classification using episodic training and task augmentation.
To address the challenge of limited annotated source domains in clinical practice and mitigate task-level overfitting.
To improve the generalization capacity of machine learning models in medical imaging despite domain variations.

Main Methods:

Developed an episodic training paradigm based on meta-learning to simulate knowledge transfer from training tasks to real DG testing tasks.
Introduced task augmentation to increase training task variety, alleviating overfitting caused by a limited number of source domains.
Incorporated a novel meta-objective to regularize deep embeddings of training domains within the learning framework.

Main Results:

The proposed DG scheme demonstrated effectiveness in experiments involving histopathological images and abdominal CT images.
Task augmentation successfully enhanced the variety of training tasks, reducing overfitting risks associated with limited source domains.
The meta-objective effectively regularized deep embeddings, contributing to improved generalization performance.

Conclusions:

The novel DG scheme, combining episodic training and task augmentation, offers a promising solution for medical imaging classification.
The method effectively addresses the challenge of limited source domains and task-level overfitting in medical DG.
The approach shows potential for improving the reliability and generalizability of machine learning models in diverse clinical settings.