[Paper Review] Joint Left Atrial Segmentation and Scar Quantification Based on a DNN with Spatial Encoding and Shape Attention

Lei Li, Xin Weng+2 more|arXiv (Cornell University)|Jun 23, 2020

Atrial Fibrillation Management and Outcomes12 references4 citations

TL;DR

This paper proposes an end-to-end deep learning framework, MTL-SESA, that jointly segments the left atrium (LA) and quantifies fibrotic scars in late gadolinium enhancement MRI. By integrating spatial encoding loss via signed distance transform maps and shape attention through surface projection, the method reduces segmentation noise and improves detection of small, discrete scars, achieving superior performance over baseline and state-of-the-art methods on 60 patient scans.

ABSTRACT

We propose an end-to-end deep neural network (DNN) which can simultaneously segment the left atrial (LA) cavity and quantify LA scars. The framework incorporates the continuous spatial information of the target by introducing a spatially encoded (SE) loss based on the distance transform map. Compared to conventional binary label based loss, the proposed SE loss can reduce noisy patches in the resulting segmentation, which is commonly seen for deep learning-based methods. To fully utilize the inherent spatial relationship between LA and LA scars, we further propose a shape attention (SA) mechanism through an explicit surface projection to build an end-to-end-trainable model. Specifically, the SA scheme is embedded into a two-task network to perform the joint LA segmentation and scar quantification. Moreover, the proposed method can alleviate the severe class-imbalance problem when detecting small and discrete targets like scars. We evaluated the proposed framework on 60 LGE MRI data from the MICCAI2018 LA challenge. For LA segmentation, the proposed method reduced the mean Hausdorff distance from 36.4 mm to 20.0 mm compared to the 3D basic U-Net using the binary cross-entropy loss. For scar quantification, the method was compared with the results or algorithms reported in the literature and demonstrated better performance.

Motivation & Objective

• To address the challenge of noisy, outlier-prone segmentation in deep learning-based left atrial (LA) and scar segmentation.
• To improve scar quantification accuracy, especially for small and discrete fibrotic regions, by leveraging spatial and shape priors.
• To develop an end-to-end trainable multi-task network that jointly optimizes LA segmentation and scar detection.
• To reduce reliance on manual initialization or post-processing steps common in conventional scar quantification methods.
• To mitigate class imbalance in scar detection by incorporating spatial and structural constraints into the loss function.

Proposed method

• Introduces a spatially encoded (SE) loss based on the signed distance transform map (DTM) to regularize LA segmentation, reducing boundary blurriness and noisy patches.
• Applies the SE loss to both LA and scar prediction tasks using distance probability maps, enhancing spatial consistency in predictions.
• Employs a shape attention (SA) mechanism that projects the LA boundary as a spatial mask to guide scar prediction, explicitly modeling the anatomical relationship between LA and scars.
• Designs a multi-task learning (MTL) U-Net architecture with two decoders—one for LA segmentation and one for scar quantification—enabling end-to-end training.
• Uses the signed DTM defined as φ(xi) = -d^β for inside the LA, 0 on the boundary, and d^β for outside, with β=1 in experiments.
• Implements a surface projection strategy to embed the SA module, ensuring that scar predictions are spatially constrained to the LA wall surface.

Experimental results

Research questions

• RQ1 Can a spatially encoded loss based on the distance transform map improve the quality of deep learning-based left atrial segmentation by reducing noisy and disconnected regions?
• RQ2 How does integrating shape attention via surface projection enhance the detection of small and discrete left atrial scars?
• RQ3 Does the joint multi-task learning framework with SE and SA losses outperform conventional loss functions (e.g., BCE, Dice) in both LA segmentation and scar quantification?
• RQ4 To what extent does the proposed method reduce the impact of class imbalance in scar detection compared to baseline deep learning approaches?
• RQ5 Can the end-to-end framework eliminate the need for manual initialization or separate post-processing steps in scar quantification?

Key findings

• The proposed MTL-SESA method reduced the mean Hausdorff distance for LA segmentation from 36.4 mm (3D U-Net with BCE loss) to 20.0 mm, significantly improving boundary accuracy.
• The SE loss improved segmentation quality by reducing noisy patches and outliers, with a statistically significant improvement in Hausdorff distance (p < 0.001), even when Dice scores were similar.
• The method achieved a Dice score of 0.867 ± 0.032 for scar quantification, outperforming all compared methods including Otsu, MGMM, LearnGC, and U-Net with BCE or Dice loss (p ≤ 0.001).
• The inclusion of both SE and SA losses led to better detection of small and discrete scars, as visually confirmed in 3D reconstructions, with smoother and more anatomically plausible scar maps.
• The method demonstrated performance comparable to inter-observer variability (Dice: 0.891 ± 0.017 for inter-observer), indicating high reliability.
• While Accuracy and G-Dice showed slight degradation with SE/SA, the overall improvement in Dice and Hausdorff distance confirms the method’s robustness and clinical relevance.