CauAug_DCMIS

The official implementation of our paper "Causality-Adjusted Data Augmentation for Domain Continual Medical Image Segmentation", published in IEEE Journal of Biomedical and Health Informatics (JBHI), 2025.

CauAug is the second work in our domain continual medical image segmentation series. Building upon TED, CauAug studies continual segmentation from a causal learning perspective and focuses on a new question:

Can both old and new knowledge become biased during continual learning because of spurious correlations and domain-specific confounders?

To address this problem, CauAug introduces causality-adjusted data augmentation for jointly optimizing:

• old knowledge retention;
• new knowledge learning;
• continual adaptation under domain shifts.

📖 Table of Contents

• CauAug_DCMIS
  • 📖 Table of Contents
  • Overview
  • What's New
  • Research Motivation
  • Method Evolution
  • Key Idea
  • Framework
  • Requirements
  • Project Structure
  • Data Preparation
  • Run
  • Analysis
  • Ablation Study
  • Acknowledgement
  • Citation

Overview

Domain continual medical image segmentation aims to continuously adapt segmentation models to sequentially arriving medical domains without accessing previous-domain data.

Compared with traditional continual learning scenarios, medical domain shifts are often caused by:

• different imaging devices;
• different acquisition protocols;
• different hospitals and populations;
• texture and appearance variations.

Although knowledge distillation can alleviate catastrophic forgetting, existing methods mainly focus on preserving old knowledge and overlook an important issue:

Both old and new knowledge may simultaneously contain causal bias.

These biases are introduced by spurious correlations and domain-specific confounders, which may cause the segmentation model to rely on irrelevant local textures or unstable domain appearance instead of true anatomical structures.

CauAug addresses this problem by introducing causal intervention mechanisms into domain continual segmentation.

What's New

This work extends our previous TED framework from a causal perspective.

• TED_DCMIS focuses on enhancing old knowledge learning through tri-enhanced distillation.
  • 📄 Paper (MedIA)
  • 💻 Code (GitHub)

Our subsequent work further extends CauAug:

• TKRL_DCMIS focuses on teacher-originated defects and rectifies inherited knowledge gaps and biases inside old teacher models.
  • 📄 Paper (IEEE JBHI)
  • 💻 Code (GitHub)

Research Motivation

TED improves old knowledge retention by enhancing:

• diversity;
• transfer accuracy;
• fusion stability.

However, continual segmentation still suffers from a deeper issue.

In domain continual medical image segmentation, the segmentation model may learn spurious correlations caused by:

• irrelevant local textures;
• domain-specific appearance features.

As a result:

1. New knowledge learning becomes biased

   • the model may overfit texture-level confounders;
   • anatomical generalization becomes weak.

2. Old knowledge retention becomes biased

   • domain-specific information may dominate knowledge distillation;
   • old knowledge may not be comprehensively preserved.

Therefore, CauAug studies continual segmentation from a causal perspective and attempts to jointly optimize both old and new knowledge learning under domain shifts.

Method Evolution

CauAug is the second stage of our research line on domain continual medical image segmentation.

TED
└── How to better retain old knowledge?
    ├── diversity enhancement
    ├── transfer accuracy enhancement
    └── fusion stability enhancement

CauAug
└── How to causally optimize both old and new knowledge?
    ├── causal intervention
    ├── causal augmentation
    └── confounder disentanglement

The overall evolution of this research series is:

TED: Old knowledge retention
  ↓
CauAug: Causal learning of both old and new knowledge
  ↓
TKRL: Rectification of teacher-originated defects

Compared with TED, CauAug no longer treats continual learning as only a knowledge retention problem. Instead, it asks whether the retained knowledge itself may already contain causal bias caused by domain-specific confounders.

Key Idea

The core idea of CauAug is:

Continual segmentation should learn causal anatomical representations instead of domain-specific spurious correlations.

To achieve this goal, CauAug introduces causality-adjusted augmentation mechanisms for both:

• new knowledge generalization;
• old knowledge distillation.

Specifically, CauAug introduces:

• Texture-Domain Adjustment Hybrid-Scheme (TDAHS)

  • establishes a domain-continual causal model;
  • identifies confounders in continual segmentation.

• Cross Kernel Network (CKNet)

  • removes local texture confounders;
  • improves anatomical generalization for new knowledge learning.

• Fourier Transformer Generator (FTGen)

  • selectively restores old domain-specific features;
  • improves comprehensive old knowledge distillation.

Framework

CauAug introduces a causal intervention framework for domain continual segmentation.

The framework contains two causal pathways:

1. Anatomy Causality

   • focuses on anatomical structures;
   • removes irrelevant local texture confounders.

2. Domain Causality

   • focuses on domain-specific features;
   • restores representative old-domain knowledge.

The overall optimization objective jointly combines:

• segmentation learning for the current domain;
• causal new knowledge learning;
• causal old knowledge distillation.

Requirements

• Python 3.8.15
• PyTorch
• CUDA

Install dependencies:

pip install -r requirements.txt

Project Structure

--ablation/
--ablation_results/
--analysis/
--data_prep/
--mp/
--storage/
--README.md
--requirements.txt
--main.py
--get.py
--args.py
--command

Data Preparation

Please refer to the data preparation instructions:

cat data_prep/readme.md
python data_prep/prostate_prepare.py
python data_prep/cardiacmm_prepare.py
python data_prep/optic_prepare.py
python data_prep/hippocampus_prepare.py

Run

Please check the example commands:

cat command

Example for prostate continual segmentation:

python main.py --dataset prostate --approach cauaug --epochs 50 \
  --experiment-name prostate-cauaug-unet --backbone unet --device-ids 5 \
  --DomAug --AnaAug

Analysis

python analysis/eval_dataset.py   # evaluate each dataset and each approach
python analysis/table_figure.py   # generate tables and figures in the paper
python analysis/save_images.py    # save segmentation results
python analysis/effi.py           # generate dynamic continual curves

Ablation Study

# Ablation study of CKNet for anatomy causality
python ablation/ana_ablation.py

# Ablation study of FTGen for domain causality
python ablation/dom_ablation.py

# Draw continual-learning tables and curves
ablation/abation.ipynb

Acknowledgement

Our code is inspired by ACS and our previous work TED.

Citation

@ARTICLE{11054328,
  author={Zhu, Zhanshi and Dong, Qing and Luo, Gongning and Wang, Wei and Dong, Suyu and Wang, Kuanquan and Tian, Ye and Wang, Guohua and Li, Shuo},
  journal={IEEE Journal of Biomedical and Health Informatics}, 
  title={Causality-Adjusted Data Augmentation for Domain Continual Medical Image Segmentation}, 
  year={2026},
  volume={30},
  number={2},
  pages={1429-1442},
  keywords={Image segmentation;Biomedical imaging;Anatomical structure;Data augmentation;Data models;Bioinformatics;Adaptation models;Absorption;Electronic mail;Continuing education;Causal data augmentation;domain continual learning;knowledge distillation;medical segmentation},
  doi={10.1109/JBHI.2025.3584068}}