Deep learning-based cardiac cine segmentation: Transfer learning application to 7T ultrahigh-field MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Artificial neural networks (ANNs) excel at cardiac MRI segmentation but require extensive annotated data.
• Generalization of ANNs is limited across different MRI vendors, field strengths, and pathologies.
• Transfer learning offers a solution, but data requirements remain unclear.

Purpose of the Study:

• To assess data requirements for transfer learning in challenging 7T cardiac MRI segmentation.
• To provide guidelines, tools, and annotated data for transfer learning in cardiac MRI.

Main Methods:

• A segmentation model was trained on public cardiac cine MRI data.
• Transfer learning was applied to 7T cine MRI data from healthy volunteers.
• Data subsets of varying sizes and structures were used to evaluate transfer learning efficacy.

Main Results:

• Pre-trained models achieved DICE scores of 0.835 (LV) and 0.670 (My).
• Transfer learning improved scores to 0.900 (LV) and 0.791 (My).
• Using only end-systolic/diastolic images reduced data by 90% without performance loss.

Conclusions:

• Transfer learning demonstrably enhances cardiac cine image segmentation.
• Guidelines and resources are provided to facilitate transfer learning adoption.
• Publicly available data, models, and code support future research.