From Scarcity to Synthesis: Continual Learning Integrates Supervised and Unsupervised CT Image Recovery Models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for CT image recovery requires large datasets, but data scarcity and fragmentation across different scanner types and training paradigms pose significant challenges.
• Continual learning (CL) offers a solution by enabling sequential training across datasets, but it often suffers from catastrophic forgetting, where prior knowledge is lost.

Purpose of the Study:

• To investigate the relationship between catastrophic forgetting and model performance in CT image recovery.
• To propose a novel task-agnostic continual learning framework to address data scarcity and catastrophic forgetting in CT image recovery.

Main Methods:

• Extensive empirical analysis to observe the correlation between forgetting and performance in CL for CT image recovery.
• Development of a task-agnostic CL framework utilizing selective experience replay, balancing supervised and unsupervised tasks.
• Implementation of dual-weight knowledge distillation to preserve past predictions while learning new tasks.

Main Results:

• Catastrophic forgetting in CT image recovery inversely correlates with model performance on new datasets.
• Shared, transferable representations and knowledge exist across different CT datasets and training paradigms.
• The proposed framework significantly reduced forgetting and enhanced knowledge transfer compared to state-of-the-art baselines, approaching the joint training upper bound.

Conclusions:

• Continual learning is a practical solution for CT image recovery, effectively enabling cross-paradigm knowledge sharing.
• The proposed framework achieves strong reconstruction fidelity and sustained performance across diverse CT datasets.
• Reduced forgetting in CL is crucial for better retention and adaptability in CT image recovery tasks.