Multi-Centre, Multi-Vendor and Multi-Disease Cardiac Segmentation: The M&Ms Challenge.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Cardiovascular imaging

Background:

• Deep learning has significantly improved cardiac magnetic resonance (CMR) segmentation accuracy.
• Current models often lack generalizability due to training on limited, homogeneous datasets.
• This limits their application across different clinical settings, vendors, and imaging protocols.

Purpose of the Study:

• To address the generalizability limitations of deep learning models for CMR segmentation.
• To present the findings of the Multi-Centre, Multi-Vendor and Multi-Disease Cardiac Segmentation (M&Ms) Challenge.
• To foster research and benchmarking in generalizable deep learning for cardiac segmentation.

Main Methods:

• Organized the M&Ms Challenge at MICCAI 2020, involving 14 participating teams.
• Teams employed diverse deep learning models, data augmentation, and domain adaptation techniques.
• Utilized a newly released, heterogeneous dataset of 375 CMR scans from multiple vendors, countries, and hospitals.

Main Results:

• Intensity-driven data augmentation proved crucial for improving model performance.
• Significant challenges remain in achieving generalizability across unseen scanner vendors and imaging protocols.
• The challenge highlighted the effectiveness of various approaches but underscored the need for further innovation.

Conclusions:

• Generalizability remains a key challenge for deep learning-based CMR segmentation.
• Intensity-driven data augmentation is important, but not sufficient for robust performance.
• The open-access dataset will facilitate future research into developing more robust and generalizable cardiac segmentation models.