Automatic 3D Surface Reconstruction of the Left Atrium From Clinically Mapped Point Clouds Using Convolutional Neural...

Area of Science:

• Medical Imaging
• Computational Anatomy
• Artificial Intelligence

Background:

• Point clouds are efficient data formats for anatomical information.
• Current left atrium (LA) visualization from clinical mapping point clouds is limited and difficult to validate.
• Additional imaging modalities like MRI/CT are often required to enhance LA mapping accuracy.

Purpose of the Study:

• To propose a novel deep learning framework for automatic 3D surface reconstruction of the left atrium directly from clinical mapping point clouds.
• To address limitations in current LA visualization methods used during cardiac ablation procedures.
• To provide a more efficient and cost-effective approach for 3D LA reconstruction.

Main Methods:

• A 30-layer 3D fully convolutional neural network (CNN) forms the framework's backbone.
• The CNN architecture incorporates skip connections for multi-resolution processing and large kernels to handle sparse point cloud data.
• Residual blocks and activation normalization were implemented to enhance feature learning from sparse inputs.

Main Results:

• The proposed deep learning framework achieved automatic 3D left atrium reconstruction from point clouds.
• The model demonstrated high accuracy with excellent dice scores of 93% and surface-to-surface distances below 1 pixel on independent clinical datasets.
• The lightweight CNN design processed each patient's data in approximately 10 seconds.

Conclusions:

• This study presents the first deep learning framework for direct 3D LA reconstruction from clinical mapping point clouds.
• The framework offers an efficient, cost-effective solution for 3D LA reconstruction during ablation procedures.
• The approach has the potential to improve the treatment of cardiac diseases through enhanced visualization and mapping.