Automatic Bi-Atrial Segmentation and Biomarker Extraction from Late Gadolinium-Enhanced MRI Using Deep Learning
Fan Feng1, James Kennelly1, Zhaohan Xiong1
1Auckland Bioengineering Institute, The University of Auckland, Auckland, New Zealand.
Expert Systems with Applications
|February 16, 2026
Summary
A new deep learning tool, biAtriaNet, accurately segments both atria and quantifies fibrosis, atrial wall thickness, and chamber volumes from LGE-MRIs. This advances personalized atrial fibrillation ablation strategies.
Area of Science:
- Cardiovascular Imaging
- Artificial Intelligence in Medicine
- Biomedical Engineering
Background:
- Atrial fibrillation (AF) involves progressive atrial remodeling, including dilation and fibrosis, impacting treatment efficacy.
- Late gadolinium-enhanced (LGE) MRI quantifies left atrium (LA) fibrosis but lacks robust segmentation for both atria and accurate biomarker assessment.
- Current methods often exclude the right atrium (RA) and struggle with precise anatomical and fibrotic characterization.
Purpose of the Study:
- To introduce biAtriaNet, a deep learning pipeline for automated segmentation and biomarker extraction from LGE-MRIs of both LA and RA.
- To evaluate atrial fibrosis, atrial wall thickness (AWT), and chamber dimensions/volumes for improved AF ablation guidance.
- To develop a robust tool for patient-specific AF treatment strategies.
Main Methods:
- Developed biAtriaNet, a deep learning pipeline using two CNNs with a modified U-Net architecture, residual connections, and batch normalization.
- Trained and validated on 2D cine-MRIs (UK Biobank, n=4860) and 3D LGE-MRIs (University of Utah, n=60).
- Independently tested on 11 3D LGE-MRIs (Waikato Hospital, New Zealand), comparing against expert annotations and ground truth.
Main Results:
- biAtriaNet achieved high segmentation accuracy (Dice scores: LA 91.1%, RA 88.6%) and transferability to independent datasets.
- Chamber volume and AWT measurements demonstrated high accuracy (>90% and 95.9% for LA, 94.6% for RA, respectively).
- Fibrosis estimates showed strong correlations (Kolmogorov-Smirnov: LA 86.3%, RA 90.6%, p < 0.05).
Conclusions:
- biAtriaNet enables accurate, automated, bi-atrial segmentation and biomarker extraction from LGE-MRIs.
- The pipeline provides reliable quantification of atrial anatomy and fibrosis, crucial for AF management.
- This tool holds significant potential for enhancing patient-specific AF ablation strategies and improving clinical outcomes.
