Privacy-Aware Continual Self-Supervised Learning on Multi-Window Chest Computed Tomography for Domain-Shift...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image analysis faces challenges due to limited annotated data and domain shifts, particularly in chest CT scans with varying window settings.
• Existing continual self-supervised learning (CSSL) methods often require reusing past data, which is impractical due to data privacy concerns.

Purpose of the Study:

• To develop a novel CSSL framework for robust feature learning from multi-window chest CT images while ensuring data privacy.
• To address catastrophic forgetting and domain shifts in medical image diagnosis through innovative continual pretraining strategies.

Main Methods:

• Implemented a latent replay-based mechanism within CSSL to mitigate forgetting during continual pretraining and maintain data privacy.
• Introduced a feature distillation technique combining Wasserstein distance-based knowledge distillation and batch-knowledge ensemble for robust representation learning.
• Utilized unlabeled chest CT images across different window settings for continual pretraining.

Main Results:

• The proposed framework demonstrated superior performance in learning meaningful, domain-shift-robust representations from chest CT images.
• Effectively captured relationships between previously learned knowledge and new information across training stages, mitigating domain shift impacts.
• Validated the approach on multi-window chest CT data, outperforming existing methods.

Conclusions:

• The novel CSSL framework offers a privacy-preserving solution for robust medical image analysis, particularly for chest CT scans with varying window settings.
• The integration of latent replay and feature distillation enhances model generalization and resilience to domain shifts.
• This approach advances the development of reliable AI models in healthcare by addressing data scarcity and privacy constraints.