Self-supervised learning for multi-center magnetic resonance imaging harmonization without traveling phantoms

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiomics

Background:

• Multi-center magnetic resonance imaging (MRI) datasets are crucial for advancing artificial intelligence (AI) in medical diagnosis.
• Variability in MRI data across different centers due to scanner and parameter differences negatively impacts AI model generalization.
• Existing harmonization methods often require traveling phantoms, which are impractical for large-scale datasets.

Purpose of the Study:

• To propose a novel self-supervised harmonization (SSH) method for improving the performance of AI-based diagnostic models using multi-center MRI data.
• To address the challenge of inter-center variability in MRI datasets without the need for traveling phantoms.
• To evaluate the effectiveness of the SSH method in enhancing image fidelity, reducing inter-center differences, and improving downstream model performance.

Main Methods:

• The proposed self-supervised harmonization (SSH) method treats image harmonization as an unpaired image-to-image translation problem.
• A two-stage transformation process is employed, combining a modified cycle generative adversarial network (cycleGAN) for style transfer and a histogram matching module for structure fidelity.
• The method was validated using female pelvic MRI images from two 3 Tesla (T) systems and compared against state-of-the-art and conventional techniques.

Main Results:

• The SSH method demonstrated improved image sharpness and structure fidelity compared to existing methods.
• It significantly reduced the number of radiomics features with significant inter-center differences (from 64 to 45), outperforming dualGAN, cycleGAN, ComBat, and CLAHE.
• In downstream cervical cancer classification, the SSH method achieved a higher area under the receiver operating characteristic curve (0.894) compared to other methods.

Conclusions:

• The proposed self-supervised harmonization (SSH) method effectively reduces inter-center differences in radiomics features, enhancing the generality of AI diagnostic models.
• SSH achieves superior image fidelity and significantly improves the performance of downstream tasks, such as cervical cancer classification.
• This approach offers a practical solution for harmonizing multi-center MRI data without relying on traveling phantoms.