Causality-Adjusted Data Augmentation for Domain Continual Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Distillation-based methods in domain continual medical image segmentation aim to prevent catastrophic forgetting.
• Existing methods often suffer from biases towards both new and old knowledge due to confounding factors, impacting segmentation performance.

Purpose of the Study:

• To introduce a novel framework, Causality-Adjusted Data Augmentation (CauAug), to mitigate biases in domain continual medical image segmentation.
• To improve model generalization and knowledge retention during continual learning.

Main Methods:

• Proposed the Causality-Adjusted Data Augmentation (CauAug) framework.
• Introduced the Texture-Domain Adjustment Hybrid-Scheme (TDAHS) for causal intervention.
• Developed two causality-targeted data augmentation approaches: Cross Kernel Network (CKNet) and Fourier Transformer Generator (FTGen).

Main Results:

• CauAug effectively identifies and addresses knowledge biases caused by irrelevant local textures (L) and domain-specific features (D).
• CKNet reduces reliance on local textures, enhancing focus on anatomical structures and improving generalization.
• FTGen restores domain-specific features, aiding in comprehensive distillation of old knowledge.
• Experimental results demonstrate significant mitigation of catastrophic forgetting and superior performance over existing methods.

Conclusions:

• The CauAug framework offers a robust solution for domain continual medical image segmentation by addressing inherent knowledge biases.
• CauAug enhances model performance by improving generalization and preserving old knowledge effectively.
• The proposed methods, TDAHS, CKNet, and FTGen, contribute to advancing causal inference in medical image segmentation.