AtrialJSQnet

AtrialJSQnet performs end-to-end multi-task left atrial (LA) segmentation and atrial scar projection and quantification from late gadolinium enhanced magnetic resonance imaging (LGE MRI) to enable simultaneous delineation of LA anatomy and surface-based scar measurement.

Key Features:

• End-to-End Multi-Task Framework: Integrates LA segmentation, scar projection onto the LA surface, and scar quantification into a unified network that leverages the spatial relationship between the LA cavity and atrial scars.
• Shape Attention (SA) Mechanism: Implements a shape attention mechanism using implicit surface projection embedded in the multi-task architecture to exploit the correlation between LA anatomy and scars.
• Spatial Encoding (SE) Loss: Introduces a spatial encoding loss that incorporates continuous spatial information to reduce noise in predicted patches and improve segmentation precision.

Scientific Applications:

• Post-ablation LGE MRI analysis: Quantifies atrial scar and segments LA anatomy in post-ablation LGE MRI to assess ablation outcomes by distinguishing healthy and scarred tissue.
• Cardiac imaging research and clinical assessment: Supports research and clinical evaluation of left atrial structure and fibrosis through simultaneous segmentation and scar quantification.

Methodology:

Evaluated via experiments on public datasets, including 60 post-ablation LGE MRIs from the MICCAI2018 Atrial Segmentation Challenge and additional sets of 40 pre-ablation and 30 multi-center post-ablation LGE MRIs from the ISBI2012 Left Atrium Fibrosis and Scar Segmentation Challenge to assess domain generalization.