[Paper Review] AtrialGeneral: Domain Generalization for Left Atrial Segmentation of Multi-Center LGE MRIs

Lei Li, Veronika A. Zimmer+2 more|arXiv (Cornell University)|Jun 16, 2021

Cardiac Valve Diseases and Treatments17 references4 citations

TL;DR

This paper proposes AtrialGeneral, a domain generalization framework for left atrial (LA) segmentation in multi-center late gadolinium-enhanced MRI (LGE-MRI), addressing performance degradation due to domain shift across scanners and sites. It evaluates three domain generalization strategies—histogram matching, mutual information-based disentangled representation, and random style transfer—finding that simple histogram matching yields the best performance, improving Dice score from 0.678 to 0.772 on unseen domains.

ABSTRACT

Left atrial (LA) segmentation from late gadolinium enhanced magnetic resonance imaging (LGE MRI) is a crucial step needed for planning the treatment of atrial fibrillation. However, automatic LA segmentation from LGE MRI is still challenging, due to the poor image quality, high variability in LA shapes, and unclear LA boundary. Though deep learning-based methods can provide promising LA segmentation results, they often generalize poorly to unseen domains, such as data from different scanners and/or sites. In this work, we collect 210 LGE MRIs from different centers with different levels of image quality. To evaluate the domain generalization ability of models on the LA segmentation task, we employ four commonly used semantic segmentation networks for the LA segmentation from multi-center LGE MRIs. Besides, we investigate three domain generalization strategies, i.e., histogram matching, mutual information based disentangled representation, and random style transfer, where a simple histogram matching is proved to be most effective.

Motivation & Objective

• To evaluate the domain generalization performance of deep learning models on multi-center LGE-MRI for left atrial segmentation.
• To investigate whether domain generalization strategies can mitigate performance degradation caused by domain shift across different scanners and imaging centers.
• To compare the effectiveness of three domain generalization techniques—histogram matching, mutual information-based disentangled representation, and random style transfer—on LA segmentation in LGE-MRI.
• To analyze differences in segmentation performance between pre- and post-ablation LGE-MRI scans.
• To identify the most effective and practical domain generalization method for clinical deployment in multi-center LGE-MRI applications.

Proposed method

• The study employs four standard U-Net-based segmentation models: U-Net, U-Net++, DeepLab v3+, and MAnet, all trained on multi-center LGE-MRI data.
• Domain generalization is applied during training using three strategies: histogram matching (HM), mutual information-based disentangled representation (MID), and random style transfer (RST).
• Histogram matching aligns the intensity distribution of target domain images to that of source domains, reducing domain shift without model retraining.
• Mutual information-based disentangled representation aims to learn domain-invariant features by separating content and domain-specific representations.
• Random style transfer applies data augmentation by randomly transferring style statistics from source domains to target images during training.
• The models are evaluated using Dice score, average surface distance (ASD), and Hausdorff distance (HD) on unseen domains from different centers and scan types (pre- and post-ablation).

Experimental results

Research questions

• RQ1 How do standard deep learning segmentation models perform on left atrial segmentation when tested on unseen domains from different centers and scanners?
• RQ2 Which domain generalization strategy—histogram matching, mutual information-based disentanglement, or random style transfer—yields the best generalization performance for LA segmentation in multi-center LGE-MRI?
• RQ3 Is there a significant performance difference between pre- and post-ablation LGE-MRI scans for LA segmentation models?
• RQ4 Does the use of domain generalization strategies improve segmentation robustness across diverse imaging protocols and scanner vendors?
• RQ5 Why does a simple histogram matching outperform more complex domain generalization methods in this clinical setting?

Key findings

• All four baseline segmentation models (U-Net, U-Net++, DeepLab v3+, MAnet) showed significant performance degradation on unseen domains, with Dice scores dropping from ~0.68 to ~0.77 when domain generalization was applied.
• Histogram matching achieved the highest Dice score of 0.772 ± 0.089, significantly outperforming both MID-Net (0.741 ± 0.064) and RST-Net (0.756 ± 0.090) on the test set of unseen domains.
• Histogram matching reduced the Hausdorff distance from 42.7 ± 8.62 mm (baseline) to 26.5 ± 5.30 mm, indicating improved boundary localization on unseen data.
• The average surface distance (ASD) was reduced from 6.19 ± 1.65 mm (baseline) to 3.76 ± 1.04 mm with histogram matching, showing improved surface accuracy.
• MID-Net and RST-Net showed similar Dice performance but worse ASD and HD values compared to histogram matching, indicating less robust boundary prediction.
• Segmentation performance was more consistent on pre-ablation LGE-MRI, with lower standard deviations in Dice and HD, suggesting greater robustness compared to post-ablation scans.