Domain generalization in deep learning for contrast-enhanced imaging.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Domain generalization is crucial for deep learning in contrast-enhanced imaging due to varying clinical protocols.
• Limited multi-center contrast-enhanced data hinders the development of generalizable models.
• Need for tools to generalize single-domain, single-center models to unseen clinical settings.

Purpose of the Study:

• Evaluate deep learning techniques for generalizability in contrast-enhanced image segmentation.
• Assess methods like data augmentation, domain mixing, transfer learning, and domain adaptation.
• Demonstrate domain generalization for ventricular segmentation in contrast-enhanced cardiac MRI.

Main Methods:

• Systematic evaluation of data augmentation, domain mixing, transfer learning, and domain adaptation.
• Application to ventricular segmentation in contrast-enhanced cardiac MRI.
• Utilized a multi-center dataset from four hospitals across three countries.

Main Results:

• The combination of data augmentation and transfer learning achieved strong generalization to new clinical centers.
• Single-center models demonstrated effectiveness on unseen data from different hospitals.
• Performance comparable to or exceeding multi-center, multi-vendor models.

Conclusions:

• Single-domain deep learning models with generalization strategies can achieve robust performance in contrast-enhanced imaging.
• Eliminates the necessity for extensive multi-center datasets for training generalizable models.
• Highlights the potential of domain generalization for wider clinical adoption.