Dynamic memory to alleviate catastrophic forgetting in continual learning with medical imaging

Area of Science:

• Medical Imaging
• Machine Learning
• Artificial Intelligence in Healthcare

Background:

• Machine learning (ML) models in medical imaging are crucial for diagnosis and treatment but struggle with domain shifts caused by evolving technology and protocols.
• These shifts lead to decreased prediction accuracy and outdated models, limiting ML's clinical utility.
• Existing methods often fail to adapt to continuous changes in imaging data streams.

Purpose of the Study:

• To propose a continual learning approach to address domain shifts in medical imaging.
• To adapt ML models to new data variations without catastrophic forgetting.
• To maintain and improve prediction accuracy on diverse and evolving imaging datasets.

Main Methods:

• A continual learning framework is developed to adapt ML models to emerging variations in a continuous data stream.
• A dynamic memory mechanism is employed for data rehearsal, mitigating catastrophic forgetting.
• Pseudo-domain detection is used to balance memory and enable expansion into new data domains.

Main Results:

• The proposed continual learning method consistently demonstrated an advantage in adapting to domain shifts.
• Evaluations on cardiac segmentation (MRI) and lung nodule detection (CT) showed improved performance.
• The technique effectively counteracted catastrophic forgetting while incorporating new data variations.

Conclusions:

• Continual learning offers a robust solution for maintaining ML model performance in dynamic medical imaging environments.
• The dynamic memory and pseudo-domain detection approach successfully handles domain shifts and data evolution.
• This method enhances the reliability and longevity of ML applications in clinical practice.