A medical unsupervised domain adaptation framework based on Fourier transform image translation and multi-model...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation models struggle with heterogeneous data, leading to performance degradation.
• Existing feature-adaptation adversarial networks often face training instability.
• Robust cross-domain medical image segmentation is crucial for reliable analysis.

Purpose of the Study:

• To propose a novel unsupervised domain adaptation (UDA) framework for cross-domain medical image segmentation.
• To enhance the robustness of segmentation models when processing data with different distributions.
• To overcome the limitations of existing adversarial training methods.

Main Methods:

• Integrated Fourier transform-guided image translation with multi-model ensemble self-training.
• Utilized Fourier transform to replace source image amplitude spectra with target image spectra.
• Augmented target datasets with synthetic images and employed entropy minimization for regularization.

Main Results:

• Achieved significant improvements in segmentation accuracy on liver CT datasets.
• Dice Similarity Coefficient (DSC) increased by nearly 34% compared to models without domain alignment.
• Outperformed existing models, showing DSC improvements of 10.8% and 6.7%.

Conclusions:

• The proposed Fourier transform-based UDA framework effectively mitigates performance degradation due to domain shift.
• The method demonstrates superior performance in cross-domain segmentation tasks.
• The multi-model ensemble self-training strategy enhances the overall robustness of segmentation systems.