Deep learning for the harmonization of structural MRI scans: a survey

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Science

Background:

• Medical imaging datasets often exhibit inconsistencies due to diverse scanners, protocols, and settings across institutions.
• These variations compromise data consistency and compatibility, hindering reliable medical image processing and analysis.
• Image harmonization is essential to mitigate these effects and ensure standardized data for research.

Purpose of the Study:

• To review and analyze the latest deep learning techniques for medical image harmonization.
• To evaluate the strengths and limitations of various deep learning architectures and learning algorithms.
• To provide guidance for researchers in selecting appropriate harmonization methods.

Main Methods:

• Comprehensive overview of fundamental image harmonization strategies, datasets, metrics, and scanner characteristics.
• Analysis of deep learning harmonization techniques based on network architecture (U-Net, GANs, VAEs, transformers, etc.).
• Investigation of Disentangled Representation Learning (DRL) as a key learning algorithm.

Main Results:

• Deep learning approaches, particularly GANs and VAEs, show significant advancements in medical image harmonization.
• Different network architectures and learning strategies offer varying performance and suitability for specific harmonization tasks.
• A key limitation identified is the lack of standardized quantitative comparisons across existing methods.

Conclusions:

• Deep learning offers powerful tools for harmonizing medical imaging data, improving consistency and compatibility.
• Selecting the right architecture and learning algorithm is crucial for effective harmonization.
• Further research is needed to establish comprehensive evaluation metrics and address current limitations.