BSA-Net: Boundary-prioritized spatial adaptive network for efficient left atrial segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Atrial fibrillation necessitates precise left atrial segmentation for treatment.
• Deep learning shows promise but struggles with incomplete inputs from random cropping.
• Existing methods fail to address structural incompleteness and boundary discontinuities.

Purpose of the Study:

• To develop a novel deep learning model for robust left atrial segmentation.
• To overcome limitations of current methods regarding input data completeness and boundary accuracy.
• To improve the speed-accuracy trade-off in cardiac image analysis.

Main Methods:

• Proposed BSA-Net with an adaptive adjustment strategy for feature position and loss optimization.
• Introduced Spatial-adaptive Convolution (SConv) for cross-positional feature relationships.
• Developed dual Boundary Prioritized loss for enhanced boundary precision.

Main Results:

• Achieved Dice scores of 92.55% (LA), 91.42% (Utah), and 84.67% (Waikato).
• Demonstrated superior performance over state-of-the-art methods on benchmark datasets.
• Utilized only 2.16 M parameters, ~80% fewer than comparable models.

Conclusions:

• BSA-Net offers a significant advancement in left atrial segmentation accuracy and efficiency.
• The adaptive strategies effectively handle incomplete data and complex boundaries.
• The model provides a better speed-accuracy trade-off for clinical applications.