Topology-Preserving Deep Image Segmentation

• The paper introduces novel deep segmentation techniques that enforce topological invariants via persistent homology losses, diffeomorphic warping, and graph-based analysis.
• It integrates rigorous topological definitions like Betti numbers and Euler characteristics with deep neural networks to maintain structural accuracy in segmentation.
• These methods achieve significant reduction in topological errors and improved segmentation performance on both medical and natural scene datasets.

Topology-preserving deep image segmentation is a class of methods and frameworks for extracting object regions from images while maintaining explicit control over topological invariants—such as the number of connected components, genus (holes), or more general Betti numbers—within the predicted segmentation mask. Topologically faithful segmentations are essential in domains where pixel- or boundary-level accuracy alone is insufficient, as in neuroanatomy, vascular imaging, or materials science, where the global shape, connectivity, and containment relations are required for downstream analysis. Modern deep neural networks trained with local, per-pixel losses are prone to break structures or hallucinate spurious objects, thus violating known or desired topological properties. The topology-preserving paradigm seeks to guarantee or robustly enforce topology correctness, using tools from computational topology (persistent homology, Morse theory), geometric regularization, deformation-based approaches, graph theory, or efficient local penalization. Advances in this area now permit fast, differentiable, and scalable methods for training deep networks to produce segmentations with prescribed topology, showing consistent improvements in structural error rates and sometimes enabling strict topological guarantees.

1. Mathematical Formalisms for Topology in Image Segmentation

Topological invariants are rigorously defined via algebraic topology, especially through the Betti numbers $\beta_k$ (counting $k$-dimensional features: connected components, holes, enclosed voids), and the Euler characteristic $\chi$ (alternating sum of Betti numbers). To enable learning, continuous relaxations through persistent homology are employed, representing the evolution of features across thresholded superlevel or sublevel sets of a segmentation likelihood map. Persistent barcodes or diagrams encode the birth and death (threshold) of each connected structure or loop. Differentiable topological metrics such as the squared Wasserstein distance between persistence diagrams (Hu et al., 2019), induced Betti matchings (Stucki et al., 2022), or variants that incorporate width or spatial information (Li et al., 16 Jan 2026) are used to compare predicted and reference topologies. Alternative, computationally efficient surrogates include the Euler characteristic computed by local bit-quad or bit-octet patterns in 2D/3D (Li et al., 31 Jul 2025).

A notable approach is enforcing topology directly via diffeomorphic maps—a function $f:\Omega \to \Omega$ is a diffeomorphism if its Jacobian determinant is positive everywhere, guaranteeing preservation of topological invariants under the warp (Wyburd et al., 2021, Zhang et al., 2022, Zhang et al., 2022). In digital settings, discrete Morse theory and simple-point analysis are used to identify topology-critical points for targeted penalization (Hu et al., 2021, Hu, 2021).

2. Core Algorithmic Frameworks for Topology Preservation

Persistent Homology and Loss Functions

The foundational class of topology-aware methods compute persistent homology on segmentation likelihood maps (via cubical complexes), extract the persistence diagram in the relevant dimension ($\beta_0$, $\beta_1$), and define a loss as the minimum-cost matching (e.g., squared Wasserstein distance) between predicted and ground-truth barcodes. Critical points—pixels at which features are born or die—are identified and gradients routed through these loci (Hu et al., 2019, Hu, 2024). Methods such as TopoNet (Hu et al., 2019) and TopoLoss (Hu, 2024) demonstrate substantial reductions in Betti number errors and improved Adjusted Rand Index (ARI) and Variation of Information (VOI).

Deformable and Diffeomorphic Network Paradigms

TEDS-Net (Wyburd et al., 2021) and TPSN (Zhang et al., 2022, Zhang et al., 2022) implement segmentation by learning a dense deformation field which warps a template mask of known topology to fit the image. The field is regularized such that its Jacobian determinant remains strictly positive, ensuring that no folds, tears, or overlaps are present and thus guaranteeing preservation of Betti numbers in the segmentation. Multi-scale extensions further boost boundary fidelity while maintaining topology (Zhang et al., 2022). These frameworks consistently achieve 0% topological error rate on challenging anatomical datasets.

Graph-Based Component Analysis

Topograph (Lux et al., 2024) introduces a component graph that represents topological atoms as superpixels and encodes their adjacency. Misclassified nodes are categorized as “critical” or “regular” based on their impact on global topology. The loss aggregates over only topology-critical nodes, and by design, zero loss implies homotopy equivalence between ground truth and prediction. This approach achieves state-of-the-art topological accuracy at 3–6$\times$ faster loss computation than persistent-homology-based losses and is extensible to multiclass segmentation.

Local and Efficient Topological Approximations

Alternative surrogates rely on computationally efficient algebraic invariants, most notably the Euler characteristic $\chi$ (Li et al., 31 Jul 2025). By precomputing local $\chi$-errors patch-wise (using convolutional kernels), generating violation maps by backpropagation, and passing only these error-prone regions to a small correction network, the global Betti error is dramatically reduced at linear computational cost. Similarly, local neighbor penalization as in SCNP (Valverde et al., 19 Mar 2026) encourages correction of poor neighbor predictions and restores thin connections, improving Betti error with negligible compute and plug-and-play integration.

3. Topological Guarantees, Loss Construction, and Training Strategies

Many frameworks provide theoretical or empirical guarantees of topological faithfulness. Persistent homology-based losses guarantee correct Betti numbers when loss is zero, but this is only achieved asymptotically (Hu et al., 2019). Diffeomorphic frameworks offer strict topology preservation by construction (as long as the discretized deformation is sufficiently smooth and the boundary is mapped consistently) (Wyburd et al., 2021, Zhang et al., 2022, Zhang et al., 2022). Losses based on homotopy warping identify the minimal set of topology-critical pixels between ground truth and prediction, enforcing the minimal edit distance up to homotopy equivalence (Hu, 2021). Graph-based and Euler characteristic approaches (e.g., Topograph and FastEC) guarantee that segmentation is homotopically equivalent (or $\chi$-equivalent) to the target if the critical loss is minimized to zero (Lux et al., 2024, Li et al., 31 Jul 2025).

In all frameworks, the topology-aware loss is balanced with standard volumetric or pixel-wise losses (Dice, cross-entropy). Optimization is typically performed with Adam or SGD; in computationally intensive methods, patch-based training is employed to manage cubic or subcubic scaling of PH calculations. Some approaches offer multi-stage (coarse-to-fine) or two-stage (prediction $k$0 correction) architectures (Li et al., 31 Jul 2025, Zhang et al., 2022).

4. Extensions: Multi-class, Multi-structure, and Width-Guaranteeing Topology

Topology-aware segmentation has progressed from binary to multi-class and multi-label settings. In multi-class medical image segmentation, topological relations between classes (e.g., containment of lumen by vessel wall; exclusion of organ boundaries) are imposed by spatially targeting violated adjacency pairs using topological interaction modules (Gupta et al., 2022, Shi et al., 2023). The TI module uses convolutional neighborhood analysis to flag and penalize boundary pixels involved in forbidden containment/exclusion interactions, substantially reducing violation rates across 2D and 3D datasets. NexToU (Shi et al., 2023) further integrates graph neural network modules to encode long-range and local topological relations.

Recent advances also tackle the limitation of classical topological energies to ignore geometry; width-aware persistent homology penalizes features born/dying below a prescribed spatial scale, disallowing spurious single-pixel bridges or holes (Li et al., 16 Jan 2026). This is achieved by smoothing the maxima/minima over neighborhoods and integrating the energy over these regions, thus encoding both topological and geometric (thickness, length) requirements.

Data-augmentation–based strategies (e.g., CoLeTra (Valverde et al., 7 Mar 2025)) inject priors that visually broken structures should be connected, improving topological error even when label annotations are imperfect or ambiguous.

5. Empirical Validation: Datasets, Metrics, and Quantitative Outcomes

Across biomedical (ISBI12, ISBI13, CREMI, DRIVE, KiTS21, ACDC, BTCV, dHCP, TopCoW) and natural-scene datasets (RoadTracer, Buildings, Massachusetts Roads, Crack500), topology-preserving methods yield significant improvements in topological metrics:

• Betti-number error and Betti-matching error are reduced by 30–50% over baselines (Stucki et al., 2022, Hu, 2024, Li et al., 31 Jul 2025).
• ARI and VOI metrics consistently improve, indicating better preservation of connected components and correct partitioning (Hu et al., 2019, Hu, 2024).
• Explicit diffeomorphic schemes achieve 0% topology error on all test cases (Wyburd et al., 2021, Zhang et al., 2022, Zhang et al., 2022).
• Euler characteristic and graph-based methods maintain or exceed Dice scores while dramatically lowering topological errors (Li et al., 31 Jul 2025, Lux et al., 2024).
• Topological-interaction modules reduce violation rates by up to 10$k$1 in multiclass exclusion/containment tasks (Gupta et al., 2022, Shi et al., 2023).
• Efficient local penalties (SCNP, FastEC) match or outperform persistent-homology–based approaches for a fraction of computational cost (Valverde et al., 19 Mar 2026, Li et al., 31 Jul 2025).

Qualitative results illustrate restored connectivity in thin structures, removal of spurious artifacts, and adherence to anatomical priors, validated on both 2D and 3D tasks.

6. Limitations, Trade-offs, and Current Directions

Trade-offs persist among computational cost, fidelity, and flexibility. Persistent homology–based losses are computationally intensive, scaling cubically in patch size, and patch-based training may miss global topology on large images (Hu et al., 2019). Diffeomorphic approaches can over-smooth or thicken structures, and may exhibit degraded boundary accuracy when target shapes are highly concave or require complex warps (Wyburd et al., 2021, Zhang et al., 2022). Graph, heuristic, or $k$2-based losses may not fully substitute for PH in highly intricate structures, and multi-class generalization often requires careful design of adjacency logic (Lux et al., 2024, Li et al., 31 Jul 2025, Gupta et al., 2022).

Label noise, low contrast, and annotation errors can hamper the effectiveness of topological learning (especially when loss terms reinforce mistakes in training masks) (Valverde et al., 7 Mar 2025, Li et al., 31 Jul 2025). Parameter tuning (e.g., loss weights, neighborhood size, width thresholds) remains nontrivial, with performance sometimes sensitive to hyperparameter choices.

Active research areas include differentiable PH and DMT modules for scalable end-to-end learning, width- and curvature-aware topology constraints, multi-scale and hybrid approaches, and adaptive or self-supervised topology correction strategies.

7. Conclusion and Perspectives

Topology-preserving deep image segmentation has matured into a field offering a continuum of methods—from theoretically strict diffeomorphic warps, through persistent homology–driven supervision, efficient local surrogates, and plug-and-play neighbor penalization, to task-tailored augmentation. These frameworks yield order-of-magnitude reductions in topological error, strict homotopy guarantees in critical applications, and robust improvements in both 2D and 3D object segmentation. Ongoing developments are focused on joint geometry–topology guarantees, improved computational efficiency (patchless, multi-GPU, or GPU-PH), and more expressive, domain-aware invariants suitable for complex or multi-label anatomies. Comprehensive empirical validation confirms that topology preservation is essential not only for rigorous scientific analysis but increasingly feasible within practical, scalable deep learning pipelines.