[Paper Review] Multiview Two-Task Recursive Attention Model for Left Atrium and Atrial Scars Segmentation

Jun Chen, Guang Yang+10 more|arXiv (Cornell University)|Jun 12, 2018

Cardiac Valve Diseases and Treatments9 references4 citations

TL;DR

This paper proposes a novel multiview two-task recursive attention model (MVTT) that simultaneously segments the left atrium (LA) with pulmonary veins and atrial scars from single-shot LGE-CMRI scans using sequential and dilated residual learning with an attention mechanism. The method achieves state-of-the-art performance, significantly improving scar detection sensitivity and Dice scores over existing methods while eliminating the need for additional bright-blood MRI or registration.

ABSTRACT

Late Gadolinium Enhanced Cardiac MRI (LGE-CMRI) for detecting atrial scars in atrial fibrillation (AF) patients has recently emerged as a promising technique to stratify patients, guide ablation therapy and predict treatment success. Visualisation and quantification of scar tissues require a segmentation of both the left atrium (LA) and the high intensity scar regions from LGE-CMRI images. These two segmentation tasks are challenging due to the cancelling of healthy tissue signal, low signal-to-noise ratio and often limited image quality in these patients. Most approaches require manual supervision and/or a second bright-blood MRI acquisition for anatomical segmentation. Segmenting both the LA anatomy and the scar tissues automatically from a single LGE-CMRI acquisition is highly in demand. In this study, we proposed a novel fully automated multiview two-task (MVTT) recursive attention model working directly on LGE-CMRI images that combines a sequential learning and a dilated residual learning to segment the LA (including attached pulmonary veins) and delineate the atrial scars simultaneously via an innovative attention model. Compared to other state-of-the-art methods, the proposed MVTT achieves compelling improvement, enabling to generate a patient-specific anatomical and atrial scar assessment model.

Motivation & Objective

• To develop a fully automated method for simultaneous segmentation of left atrium (including pulmonary veins) and atrial scars from single LGE-CMRI scans.
• To overcome challenges in LGE-CMRI such as low signal-to-noise ratio, signal nulling of healthy tissue, and poor image quality in AF patients.
• To eliminate reliance on additional bright-blood MRI acquisitions and subsequent registration errors common in prior approaches.
• To improve sensitivity and accuracy in detecting small, hyper-enhanced scar regions using an attention mechanism.
• To enable patient-specific, reproducible, and reliable anatomical and scar assessment for guiding ablation therapy in atrial fibrillation.

Proposed method

• The MVTT framework employs multiview learning: axial view features are extracted via a sequential learning network with ConvLSTM, while sagittal and coronal views use dilated residual learning for multi-scale context capture.
• A recursive attention mechanism is integrated to enhance feature representation of small, low-contrast atrial scar regions by focusing on discriminative spatial patterns.
• The model uses a 3D U-Net-like architecture with encoder-decoder paths, where each convolutional layer is followed by ReLU and local response normalization (LRN).
• Feature maps are processed through max-pooling (encoder) and bilinear upsampling (decoder), with skip connections to preserve spatial details.
• The network is trained end-to-end using a combined loss function for both LA/PV and scar segmentation tasks, enabling joint optimization.
• The method leverages 3D LGE-MRI data reconstructed into axial, sagittal, and coronal slices to exploit complementary anatomical information.

Experimental results

Research questions

• RQ1 Can a single deep learning model accurately segment both the left atrium with pulmonary veins and atrial scars from a single LGE-CMRI scan without additional imaging?
• RQ2 How does incorporating multiview (axial, sagittal, coronal) feature learning improve segmentation performance compared to single-view approaches?
• RQ3 To what extent does an attention mechanism enhance the detection of small, low-contrast atrial scar regions in noisy LGE-CMRI data?
• RQ4 How does joint two-task learning (anatomy + scar) compare to separate training of LA/PV and scar segmentation?
• RQ5 What is the impact of architectural components such as ConvLSTM and dilated residual blocks on model robustness and generalization?

Key findings

• The MVTT achieved a Dice score of 0.9990 ± 0.0009 for left atrium and pulmonary vein segmentation, significantly outperforming the WHS method (0.905 vs. 0.859 sensitivity).
• For atrial scar segmentation, the MVTT achieved a Dice score of 0.7764 ± 0.1459, outperforming unsupervised methods (e.g., SD-based thresholding: Dice < 0.1) and U-Net baselines.
• The ablation study confirmed that multi-view sequential learning improved LA/PV segmentation over axial-only learning, with a 0.0067 increase in Dice score.
• The attention mechanism boosted scar segmentation sensitivity by 0.0272 (from 0.7491 to 0.7764) and improved Dice by 0.0273, proving its effectiveness in highlighting small scar regions.
• The simultaneous two-task learning framework yielded better results than separate training, indicating mutual supervision enhances feature learning.
• The model showed strong generalization, with high specificity (0.9997 ± 0.0002) and robust convergence, though performance was slightly reduced in pre-ablation cases with minimal native scarring.