Fully automatic segmentation and objective assessment of atrial scars for long‐standing persistent atrial fibrillation patients using late gadolinium‐enhanced <scp>MRI</scp>

Guang Yang, Xiahai Zhuang, Habib Khan, Shouvik Haldar, Eva Nyktari, Lei Li, Ricardo Wage, Xujiong Ye, Greg Slabaugh, Raad Mohiaddin, Tom Wong, Jennifer Keegan, David Firmin

TL;DR

This paper introduces a fully automatic pipeline for segmenting and assessing atrial scars in patients with long-standing persistent atrial fibrillation using late gadolinium-enhanced MRI (LGE-MRI). By automating a previously manual and time-consuming process, their method offers a more objective and efficient tool for stratifying patients and guiding ablation treatments.

What Problem Does This Paper Address?

Atrial fibrillation (AF) is a common heart rhythm disorder with significant morbidity and mortality. A key factor in AF is atrial fibrosis, which can be visualized using LGE-MRI. However, manually segmenting the heart chambers and the fibrotic scar tissue from these MRI scans is a laborious and subjective process, prone to inter-observer variability. This variability is particularly problematic given the often-poor image quality in AF patients. The paper aims to overcome these limitations by developing a fully automated system that can accurately segment both the cardiac anatomy and the atrial scar tissue, providing a more reliable and objective assessment for clinical decision-making.

Methods

The proposed pipeline consists of two main stages:

  1. Multi-Atlas Whole Heart Segmentation (MA-WHS): This stage uses a multi-atlas approach to segment the cardiac anatomy (specifically the left atrium and pulmonary veins) from a "Roadmap" MRI acquisition. This process involves:
  2. Atlas Propagation: Registering 30 pre-segmented whole-heart atlases to the target Roadmap image using a hierarchical registration strategy (global affine, local affine, and fully deformable registration).
  3. Label Fusion: Combining the results from the multiple atlases to generate a final segmentation. The authors introduce a Multiscale Patch-based Label Fusion (MSP-LF) method, which leverages information from different image scales to improve robustness against intensity variations and capture structural information.
  4. Inter-Scan Registration: The segmented Roadmap anatomy is then registered to the LGE-MRI data to define the endocardial LA boundary and PVs on the scar images.
  5. Atrial Scarring Segmentation: This stage focuses on segmenting the fibrotic scar tissue within the previously defined LA and PV geometry from the LGE-MRI data.
  6. Oversegmentation (SLIC Super-pixels): The LGE-MRI images are first oversegmented into perceptually meaningful super-pixels using the Simple Linear Iterative Clustering (SLIC) algorithm. This step helps to group pixels with similar characteristics and adhere to object boundaries.
  7. Supervised Classification (SVM): A Support Vector Machine (SVM) classifier is trained to distinguish between enhanced atrial scarring regions and non-enhanced tissues. The training data is constructed by:
  8. Manual "mouse clicks" by an expert cardiologist to identify enhanced regions.
  9. Combining these clicks with SLIC segmentation and morphological operations to define super-pixels as either enhanced or non-enhanced within the LA wall and PV.
  10. Extracting pixel intensity-based features (minimum, mean, standard deviation) from the super-pixels for SVM classification.

Datasets

The study utilized cardiac MR data acquired from 37 patients with long-standing persistent AF between 2011 and 2013. These datasets included both preablation (11 cases) and postablation (26 cases) scans. The authors mention using a "Left Atrium Segmentation Grand Challenge" dataset for building their multiatlas library, which was separate from their patient data.

Key Insights and Ideas

  • Automating a Subjective Process: The core innovation is the complete automation of atrial scar segmentation, moving away from manual, time-consuming, and observer-dependent methods. This has significant implications for clinical workflow and research reproducibility.
  • Multiscale Patch-based Label Fusion (MSP-LF): The development of MSP-LF for whole-heart segmentation is a key methodological contribution, showing improved performance over other label fusion techniques by effectively handling intensity variations.
  • SLIC Super-pixels for Scar Segmentation: Using SLIC super-pixels as the basic unit for classification in scar segmentation is an effective way to reduce computational complexity and leverage local image information.
  • Feature Selection for SVM: Identifying that simple intensity-based features (mean, standard deviation, minimum) from super-pixels are sufficient for SVM classification, rather than more complex texture or shape features, is an important finding for efficiency.
  • Clinical Relevance of Scar Assessment: The paper strongly emphasizes the clinical utility of accurate scar assessment for stratifying AF patients and guiding ablation strategies, potentially reducing the need for repeat procedures.

Key Results

  • Whole Heart Segmentation (MA-WHS): Achieved accurate delineations of cardiac anatomy with a mean Dice score of 89% compared to manual ground truth.
  • Atrial Scarring Segmentation: Demonstrated accurate segmentation of atrial scarring with a mean Dice score of 79% compared to manual ground truth.
  • Classification Performance: The SVM-based classification achieved 88% accuracy, 90% sensitivity, and 79% specificity, with an Area Under the Curve (AUC) of 0.91 from ROC analysis.
  • Speed: The fully automatic pipeline was significantly faster than manual methods, taking only 5.1 ± 0.7 seconds per patient case for segmentation, compared to 25-50 minutes for manual ground truth construction.
  • Fibrosis Extent: The automated method showed comparable results to manual ground truth in quantifying fibrosis extent (FEP) and identified significant differences between pre- and postablation cases.
  • Comparison to Other Methods: The proposed method outperformed simple thresholding and standard deviation-based methods, and was competitive with or superior to advanced unsupervised learning methods.

Limitations and Future Work

  • Limited Preablation Data: The study was limited by the number of preablation cases (11 out of 37) with sufficient image quality due to the challenges of scanning AF patients. The authors suggest that advancements in MRI acquisition strategies could improve this in the future.
  • Ground Truth Subjectivity: While efforts were made to ensure robustness, the ground truth was still based on manual segmentation (albeit with an automated starting point), which inherently carries some subjectivity. The authors acknowledge that even experienced cardiologists can miss regions of enhancement.
  • Gadolinium Contrast Agent Concerns: The paper briefly touches upon the concerns regarding gadolinium retention and adverse health effects, suggesting that macrocyclic agents (like gadobutrol used in the study) are preferred over linear agents. This highlights a broader clinical context for LGE-MRI.
  • Future Work: The authors envision integrating their pipeline into clinical routines for rapid, patient-specific LA and PV geometry modeling and objective atrial scarring assessment. Further development could focus on improving image quality in challenging AF cases and exploring alternative contrast agents or non-contrast imaging techniques.

Summary updated

Abstract

Compared with previously studied methods with manual interventions, our innovative pipeline demonstrated comparable results, but was computed fully automatically. The proposed segmentation methods allow LGE MRI to be used as an objective assessment tool for localization, visualization, and quantitation of atrial scarring and to guide ablation treatment.
BibTeX
@inproceedings{yang2018fully,
  author = {Guang Yang and Xiahai Zhuang and Habib Khan and Shouvik Haldar and Eva Nyktari and Lei Li and Ricardo Wage and Xujiong Ye and Greg Slabaugh and Raad Mohiaddin and Tom Wong and Jennifer Keegan and David Firmin},
  title = {{Fully automatic segmentation and objective assessment of atrial scars for long‐standing persistent atrial fibrillation patients using late gadolinium‐enhanced <scp>MRI</scp>}},
  booktitle = {Medical},
  year = {2018},
  doi = {10.1002/mp.12832},
}
Figure