Anatomically plausible segmentations: Explicitly preserving topology through prior deformations.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Anatomy

Background:

• Deep learning models achieve high performance in medical image segmentation.
• Traditional metrics fail to detect topological errors (e.g., holes, folds) in segmentations.
• Topological inaccuracies can negatively impact downstream clinical image processing tasks.

Purpose of the Study:

• To develop a deep learning segmentation network that preserves anatomical topology.
• To maintain competitive segmentation performance alongside topological correctness.
• To address the limitations of current state-of-the-art (SOTA) methods in handling topological errors.

Main Methods:

• Introduction of TEDS-Net, a novel segmentation network.
• Utilizing learned topology-preserving fields to deform a prior representation.
• Implementing modifications for stricter topology enforcement in the discrete domain.

Main Results:

• TEDS-Net successfully preserves anatomical topology in medical heart datasets.
• Demonstrated competitive segmentation performance against SOTA baselines.
• Generated segmentations contained no folding voxels, indicating full topology preservation for individual structures.
• Outperformed other baselines in overall scene topology preservation.

Conclusions:

• TEDS-Net effectively ensures anatomically plausible segmentations by preserving topology.
• The method addresses critical limitations of existing SOTA segmentation techniques.
• TEDS-Net offers a robust solution for medical image segmentation where topological accuracy is paramount.