SegMeshNet: Joint heart segmentation and mesh reconstruction with task-aware shared attention.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Biomedical engineering

Background:

• Accurate heart segmentation and 3D mesh reconstruction are crucial for medical diagnosis and research.
• Current methods often treat these tasks separately, limiting data utilization efficiency.

Purpose of the Study:

• To propose a novel joint learning framework, SegMeshNet, for simultaneous heart segmentation and 3D mesh reconstruction.
• To enhance feature interaction and representation for improved accuracy in both tasks.
• To develop a new loss function for superior mesh reconstruction quality.

Main Methods:

• Developed SegMeshNet, a joint learning framework integrating segmentation and reconstruction.
• Introduced a task-aware shared attention (TSA) module for cross-task feature interaction.
• Implemented a multi-scale feature fusion (MSF) module for enhanced feature representation.
• Proposed a curvature-weighted hyperbolic chamfer distance (wHCD) loss for improved reconstruction.

Main Results:

• SegMeshNet demonstrated superior performance compared to state-of-the-art methods on CT and MR datasets.
• The TSA and MSF modules effectively improved segmentation and reconstruction accuracy.
• The wHCD loss significantly enhanced mesh reconstruction quality.
• The model showed adaptability across different imaging modalities without requiring modality-specific designs.

Conclusions:

• SegMeshNet offers an efficient and accurate solution for joint heart segmentation and 3D mesh reconstruction.
• The proposed attention and feature fusion modules are key to the model's success.
• The framework's versatility makes it applicable to diverse medical imaging applications.