Lifelong nnU-Net: a framework for standardized medical continual learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning shows promise in medical image segmentation.
• Transitioning research to clinical practice requires continual learning, which is nascent in healthcare.
• Current methods often rely on static training, limiting adaptability.

Purpose of the Study:

• To present Lifelong nnU-Net, a framework for continual medical image segmentation.
• To lower the barrier for evaluating new continual learning methods in healthcare.
• To establish a reproducible benchmark for continual learning in medical segmentation.

Main Methods:

• Developed Lifelong nnU-Net, a framework built upon the nnU-Net architecture.
• Integrated modules for sequential model training and testing.
• Benchmarked five continual learning methods across three medical segmentation use cases.

Main Results:

• Demonstrated the broad applicability of the Lifelong nnU-Net framework.
• Provided a comprehensive outlook on the current state of continual learning in medical segmentation.
• Established a first reproducible benchmark for evaluating continual learning methods.

Conclusions:

• Lifelong nnU-Net facilitates the safe and effective integration of continual learning in clinical practice.
• The framework supports researchers and clinicians in evaluating and implementing dynamic deep learning models.
• This work marks a significant step towards advancing continual learning in medical AI.