Realistic image normalization for multi-Domain segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Conventional medical image normalization is dataset-specific, limiting the exploitation of cross-dataset information.
• This limitation negatively impacts the performance of downstream tasks like image segmentation.
• Existing methods fail to leverage complex joint information across multiple medical imaging datasets.

Purpose of the Study:

• To develop a novel approach for learning a common normalizing function across multiple medical image datasets.
• To improve image segmentation accuracy and realism by jointly normalizing data.
• To address the limitations of per-dataset normalization strategies in medical image analysis.

Main Methods:

• A fully automated, adversarial, and task-driven normalization approach was employed.
• The method learns a shared normalizing function across multiple datasets (iSEG, MRBrainS, ABIDE).
• Adversarial training optimizes a transfer function for realistic image generation and segmentation accuracy.

Main Results:

• Joint normalization across datasets yielded consistent, realistic normalized images.
• Improved image segmentation accuracy was observed, particularly with large intensity shifts.
• The proposed method achieved segmentation accuracy on par with state-of-the-art techniques.

Conclusions:

• Learning a common normalizing function across multiple datasets is superior to per-dataset methods.
• The adversarial, task-driven approach effectively enhances both image realism and segmentation performance.
• This strategy offers a robust solution for medical image normalization and analysis.