A Learning-based Framework for Topology-Preserving Segmentation using Quasiconformal Mappings
TL;DR
The Topology-Preserving Segmentation Network (TPSN) ensures that medical image segmentations always match the expected anatomical structure—no more "floating islands" or "random holes." By combining deep learning with the mathematical rigor of Quasiconformal mapping theory, the authors provide a framework that guarantees one-to-one (bijective) deformations and can even be trained without labeled data.
What Problem Does This Paper Address?
In medical imaging, anatomical structures follow strict morphological rules: a kidney is a single connected blob, and a vessel wall is a ring. However, standard deep learning models (like U-Net) perform pixel-wise classification. Because they treat every pixel independently, they often produce "topological errors"—segmenting a single organ into several disconnected pieces or leaving gaps where there should be solid tissue.
While some researchers use "shape priors" or regularization terms to encourage better shapes, these are often "soft" constraints that can still fail under noise. The challenge is to create a "hard" constraint within a neural network that mathematically guarantees the output mask will have the same topology as a predefined template.
Methods
The authors introduce TPSN, which moves away from pixel classification and toward deformable models. Instead of predicting "is this pixel a kidney?", the network predicts "how do I warp this template circle to fit the kidney in the image?"
The framework consists of two main parts:
- Deformation Estimation Network (DEN): An encoder-decoder architecture that takes an image and a template mask as input. It outputs a deformation map ($f$) and is regularized by a ReLU-Jacobian term. This term penalizes any "folding" in the map (where the Jacobian determinant becomes negative), promoting a one-to-one mapping.
- Beltrami Adjustment Module (BAM): This is the core innovation. Based on Quasiconformal mapping theory, every deformation has an associated Beltrami coefficient ($\mu$) that measures local geometric distortion. If $|\mu| < 1$, the mapping is guaranteed to be bijective (one-to-one). The BAM calculates $\mu$, truncates it to ensure it stays below 1, and then reconstructs a "corrected" deformation map. This provides a mathematical safety net, ensuring topological correctness even in extremely noisy images.
The authors also introduce a Fill First Dig Second (FFDS) strategy to handle complex shapes (like rings) and a multi-level (coarse-to-fine) architecture to improve accuracy on intricate boundaries.
Datasets
The model was evaluated on four diverse datasets:
- KiTS21: 3D volumetric images for kidney segmentation (simply-connected).
- ACDC: 2D cardiac images for segmenting genus-1 (ring-like) structures.
- BTCV: Multi-organ segmentation (spleen, kidneys, liver, etc.) from the MICCAI 2015 challenge.
- Ham10000: Skin lesion images used specifically to test the unsupervised version of the model.
Key Insights and Ideas
- Topology as a Hard Constraint: Unlike previous methods that use "topological loss" (which can still be ignored by the optimizer), the BAM module acts as a post-processing layer within the network that forces the mapping to be bijective.
- Unsupervised Learning: By using a Mumford-Shah-style fidelity term (which looks for regions of consistent intensity), the network can learn to segment objects without any ground-truth labels, relying entirely on the deformation of the template.
- Robustness to Noise: In experiments with "corrupted" data (blur, metal artifacts, Gaussian noise), TPSN maintained perfect topology while standard models like U-Net produced fragmented, clinically useless results.
- The Power of $\mu$: Using the Beltrami coefficient allows the network to control the "regularity" of the deformation much more effectively than standard Laplacian smoothing.
Key Results
- Topological Perfection: In almost all tests, TPSN achieved a Betti error of 1.00, meaning the number of connected components and holes in the segmentation perfectly matched the template.
- Accuracy: On the ACDC dataset, the multi-level TPSN (mlTPSN) achieved a Hausdorff Distance of 2.04, significantly outperforming TeTrIs (5.33) and TEDSNet (4.80).
- Efficiency: The unsupervised version (unTPSN) can segment an image in just 0.02 seconds, compared to 2.87 seconds for traditional optimization-based methods.
- Multi-Object Handling: The framework successfully segmented 8 different organs simultaneously in the BTCV dataset, maintaining correct topology for each.
Limitations and Future Work
The primary limitation is that the topology of the template is fixed. If you use a "blob" template to segment an organ that actually has a hole (due to pathology or unusual anatomy), the network will force it into a blob shape.
The authors suggest future work could involve:
- Integrating interactive segmentation, allowing users to modify the template topology in real-time.
- Incorporating more complex geometric features, such as shape signatures, to further refine boundary accuracy.
- Exploring convexity constraints to ensure the segmented regions aren't just topologically correct, but also anatomically plausible in their curvature.
Summary updated