CASNet: curvature-aware cardiac MRI segmentation with multi-scale and attention-driven encoding for enhanced...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate segmentation of cardiac structures in magnetic resonance imaging (MRI) is crucial for diagnosing and analyzing cardiovascular diseases.
• Conventional convolutional neural networks (CNNs) face challenges in maintaining semantic consistency and geometric smoothness in cardiac MRI segmentation, especially with anatomical variability.

Purpose of the Study:

• To introduce CASNet, a novel U-Net-based architecture designed to improve the accuracy and robustness of cardiac MRI segmentation.
• To address limitations in semantic consistency and geometric smoothness in existing segmentation models.

Main Methods:

• Proposed CASNet architecture featuring a Multi-Scale Context Block (MSCB) for enriched feature representation across scales.
• Implemented Cross-Attentive Skip Connections (CASC) for selective feature aggregation and improved feature reuse in the decoder.
• Incorporated a Curvature-Aware Loss function to enhance boundary smoothness and anatomical plausibility.

Main Results:

• CASNet demonstrated superior performance compared to baseline U-Net and attention-based models on the ACDC dataset.
• Achieved significant improvements in both region overlap and boundary accuracy metrics for cardiac structure segmentation.
• The Curvature-Aware Loss effectively improved the smoothness and anatomical plausibility of predicted segmentation boundaries.

Conclusions:

• CASNet offers a robust and generalizable solution for high-precision cardiac MRI segmentation.
• The proposed enhancements effectively address limitations in semantic consistency and geometric smoothness.
• This approach provides a strong foundation for future AI-assisted cardiac analysis and clinical applications.