Related Experiment Video
Updated: Dec 13, 2025
04:25
Author Spotlight: Bridging Gaps in Anatomy and Establishing a Foundation for Algorithmic Studies
Published on: December 15, 2023
3.5K
Improving the Generalizability of Convolutional Neural Network-Based Segmentation on CMR Images
Chen Chen1, Wenjia Bai2,3, Rhodri H Davies4,5
1Biomedical Image Analysis Group, Department of Computing, Imperial College London, London, United Kingdom.
Frontiers in Cardiovascular Medicine
|July 28, 2020
Summary
This study introduces improved data normalization and augmentation for Convolutional Neural Networks (CNNs) to enhance cardiac MRI segmentation across different scanners and sites. The method ensures accurate heart structure assessment, improving model generalizability in clinical settings.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Cardiovascular Imaging
Background:
- Convolutional Neural Networks (CNNs) offer automated cardiac MRI segmentation.
- CNN performance degrades with data from different scanners/sites.
- Generalizability is a key challenge for clinical adoption.
Purpose of the Study:
- To develop a robust CNN segmentation method for cardiac MRI.
- To improve model generalization across multi-site and multi-scanner data.
- To enable reliable heart structure and function assessment in diverse clinical settings.
Main Methods:
- Designed specialized data normalization and augmentation strategies.
- Trained a CNN on UK Biobank data (3,975 subjects).
- Validated on UK Biobank (600 subjects), ACDC (100 subjects), and BSCMR-AS (599 subjects) datasets for cross-domain testing.
Main Results:
- Achieved high segmentation accuracy comparable to literature on UK Biobank.
- Demonstrated strong cross-domain performance: Dice scores of 0.90 (LV), 0.81 (Myo), 0.82 (RV) on ACDC.
- Attained Dice scores of 0.89 (LV), 0.83 (Myo) on BSCMR-AS, showing robustness.
Conclusions:
- The proposed method significantly improves CNN generalizability for cardiac MR image segmentation.
- This approach addresses cross-scanner and cross-site variability effectively.
- Offers a practical solution for deploying AI in diverse clinical cardiac imaging environments.
