Anatomy-Aware 3D Mesh Refinement of Pericardium Segmentations on Computed Tomography

Automated pericardium segmentation in cardiac CT is notoriously difficult due to the structure's thin, low-contrast nature. While deep learning models can achieve high voxel-wise accuracy, they often produce anatomically implausible results, such as intersecting internal heart structures or failing to encapsulate the heart correctly. This paper asks whether an unsupervised, physics-inspired mesh refinement framework can use anatomical context to correct these errors post-hoc.

The authors propose a 3D iterative mesh refinement method that treats the pericardium as a deformable surface. The framework uses a two-pronged approach to guide vertex movement:

1. Anatomical Forces: A vector field is constructed using masks of surrounding organs (e.g., heart chambers, lungs). Internal structures exert an outward repulsive force, while external structures exert an inward force, pushing the mesh into the correct anatomical corridor.
2. Geometric Forces: To maintain surface integrity, the framework applies Laplacian smoothing, normal consistency, and edge length regularization. These forces ensure the mesh remains smooth and avoids artifacts like vertex clustering or sharp, non-biological creases.

The entire process is implemented in PyTorch3D as a gradient-based, GPU-accelerated optimization, allowing it to be applied as a model-agnostic post-processing step to any initial segmentation.

The method was evaluated on both high-resolution in-house data and a sparse, open-source dataset. The refinement consistently improved volumetric (Dice), surface (ASSD, HD95), and anatomical metrics across all tested scenarios. Notably, the framework showed the most significant performance gains when applied to weaker initial segmentations, such as those from models trained on limited data or applied in out-of-domain settings. This suggests the method is a robust tool for improving the reliability of pericardial masks, which is critical for downstream tasks like quantifying epicardial adipose tissue (EAT).