Progressive Learning with Anatomical Priors for Reliable Left Atrial Scar Segmentation from Late Gadolinium Enhancement MRI
Summary: arXiv:2603.26186v1 Announce Type: cross
Abstract: Cardiac MRI late gadolinium enhancement (LGE) enables non-invasive identification of left atrial (LA) scar, whose spatial distribution is strongly associated with atrial fibrillation (AF) severity and recurrence. However, automatic LA scar segmentation remains challenging due to low contrast, annotation variability, and the lack of anatomical constraints, often leading to non-reliable predictions. Accordingly, our aim was to propose a progressive learning strategy to segment LA scar from LGE images inspired from a clinical workflow.
We implemented a 3-stage framework based on SwinUNETR, which consists of the following components:
- Stage 1: A first LA cavity pre-learning model.
- Stage 2: A dual-task model that further learns spatial relationships between LA geometry and scar patterns.
- Stage 3: Fine-tuning on precise segmentation of the scar.
Furthermore, we introduced an anatomy-aware spatially weighted loss that incorporates prior clinical knowledge by constraining scar predictions to anatomically plausible LA wall regions while mitigating annotation bias. This approach is designed to enhance the accuracy of segmentation by aligning model predictions with clinical realities.
Our preliminary results were obtained on validation LGE volumes from the LASCARQS public dataset, utilizing a 5-fold cross-validation technique. The performance metrics included:
- LA segmentation Dice score: 0.94
- LA scar segmentation Dice score: 0.50
- Hausdorff Distance: 11.84 mm
- Average Surface Distance: 1.80 mm
These results demonstrate a significant improvement over a conventional one-stage scar segmentation model, which achieved:
- Dice score: 0.49
- Hausdorff Distance: 13.02 mm
- Average Surface Distance: 1.96 mm
By explicitly embedding clinical anatomical priors and diagnostic reasoning into deep learning, our proposed approach enhances the accuracy and reliability of LA scar segmentation from LGE. This research underscores the importance of clinically informed model design and its potential impact on improving patient outcomes in atrial fibrillation management.