Introduction
Atrial fibrillation (AF) is a significant contributor to global cardiovascular morbidity, often treated via radiofrequency ablation. This procedure creates scars on the left atrial (LA) wall to disrupt aberrant electrical pathways. Accurate quantification of these scars from Late Gadolinium Enhanced (LGE) MRI is crucial for procedural planning and post-operative assessment. However, automated segmentation faces severe challenges: the LA wall is exceptionally thin (1–2 mm), scars are small and sparsely distributed (class imbalance), and voxel-wise methods often produce anatomically inconsistent results.
Methodology: The MTL-SESA Net
To address these issues, researchers developed MTL-SESA Net, a multi-task learning framework designed for simultaneous LA segmentation and scar quantification. The architecture employs a shared encoder with two specialized decoders.
Spatially Encoded (SE) Loss
Standard binary cross-entropy (BCE) loss treats voxels independently, often ignoring spatial context. The MTL-SESA Net introduces a Spatially Encoded loss based on Signed Distance Transform Maps (DTM). By training the network to predict the distance of each voxel to the target boundary, the model learns the continuous geometric structure of the LA. This approach effectively penalizes isolated "noisy" predictions and enforces anatomical plausibility.
Shape Attention (SA) Mechanism
Recognizing that atrial scars are physically constrained to the LA wall, the model implements a Shape Attention mechanism. The predicted LA boundary from the first decoder serves as a spatial prior for the scar-quantification decoder. This "surface projection" focuses the network's attention on relevant anatomical regions, mitigating the impact of class imbalance and reducing false positives in the blood pool or surrounding tissue.
Experimental Results
The model was evaluated using 60 LGE MRI scans from the MICCAI 2018 LA Challenge.
Quantitative Performance
MTL-SESA Net demonstrated superior performance over baseline 3D U-Nets and state-of-the-art methods:
- LA Segmentation: The mean Hausdorff Distance (HD)—a measure of the maximum distance between prediction and ground truth—improved significantly, dropping from 36.4 mm to 20.0 mm.
- Scar Quantification: The model achieved a Dice score of 0.543, outperforming traditional Otsu thresholding (0.420) and Gaussian Mixture Models (0.499).
Qualitative Assessment
Visual inspections confirm that MTL-SESA Net produces more contiguous and anatomically accurate scar patterns compared to competing methods. The integration of spatial priors ensures that scars are correctly localized on the atrial wall.
Conclusion
By integrating spatial encoding and shape-based attention, MTL-SESA Net overcomes the inherent difficulties of thin-wall segmentation and small-target detection in cardiac MRI. The framework provides a robust, clinically viable tool for AF management, processing full 3D volumes in approximately 8 seconds. Future work aims to incorporate multi-expert consensus for ground truth and extend the model to pre-ablation fibrosis detection.
Summary updated