Stepwise incremental pretraining for integrating discriminative, restorative, and adversarial learning.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Self-supervised learning (SSL) in 3D medical imaging is complex.
• Existing SSL methods lack comprehensive integration of learning components.

Purpose of the Study:

• To develop a unified framework integrating discriminative, restorative, and adversarial self-supervised learning (SSL).
• To propose a stepwise incremental pretraining strategy for stabilizing complex SSL models in 3D medical imaging.

Main Methods:

• Developed a United framework combining discriminative, restorative, and adversarial SSL.
• Redesigned nine prominent SSL methods within the United framework.
• Implemented stepwise incremental pretraining: discriminative, then encoder-decoder, then full model training.

Main Results:

• Stepwise incremental pretraining stabilizes the United framework for 3D medical imaging.
• Achieved significant performance gains across classification and segmentation tasks.
• Demonstrated reduction in annotation costs through transfer learning.

Conclusions:

• The United framework with stepwise incremental pretraining enhances SSL performance in 3D medical imaging.
• Synergy between SSL components, enabled by stepwise pretraining, drives performance improvements.
• The approach offers a robust solution for complex medical imaging tasks, reducing data dependency.