Topological losses for segmentation and shape learning count holes, loops and connected pieces. They are built on persistent homology, a tool that tracks when each such feature appears and disappears as a scale parameter grows. Its summary of a shape, the persistence diagram, lists each feature only by those two times. It does not record where the feature sits. So when a training loss compares a predicted diagram with the ground-truth diagram by solving an optimal one-to-one assignment between their features, it may pair a hole in the top-left of the image with one in the bottom-right because the two lasted about as long. Stucki et al. found that this standard Wasserstein matching pairs features to spatially wrong partners over 99% of the time on complex benchmarks. Their fix, Betti matching, is correct and algebraic: it builds a shared comparison image, computes induced maps between persistence barcodes, and matches only features that correspond through that image. The price is steep. Worst-case cost is cubic in the number of pixels, training is limited to 48×48 patches, the 3D version (Stucki et al. 2024) still runs on the CPU, and the method needs a common pixel grid, so it does not apply to point clouds or meshes.
We propose a much simpler repair and a careful test of whether it is enough. We keep the standard optimal assignment between diagram features, solved exactly with the Hungarian algorithm (a polynomial-time solver for one-to-one matching), and change only its cost. Each feature carries a spatial footprint: a normalized mass over the pixels, voxels or points of a representative cycle. The cost of pairing two features adds the distance between their footprints to the usual difference in birth and death times. The assignment runs over features, not pixels, so it is small. Our central empirical question is whether this geometric proxy recovers the pairings of algebraic induced matching often enough to train equally good models at lower cost. If it does, the same loss carries over unchanged to point clouds and meshes, where no induced matching is available.
Core Claim
Let $p_i = (b_i, d_i, \rho_i)$ be a predicted feature with birth time $b_i$, death time $d_i$ and spatial footprint $\rho_i$, a probability measure on the domain. Let $q_j$ be a ground-truth feature defined the same way. We solve
$$\min_{M} \sum_{(i,j)\in M} \Big[\, \|(b_i,d_i)-(b_j,d_j)\|^2 + \lambda\, W_1(\rho_i,\rho_j)^2 \Big] + \sum_{k\ \text{unmatched}} \tfrac{1}{2}(d_k-b_k)^2 ,$$
where $M$ is a one-to-one partial matching and $W_1$ is the earth-mover distance between footprints. Unmatched features pay their squared distance to the diagonal, as in standard diagram distances. Setting $\lambda = 0$ recovers the spatially blind Wasserstein loss.
The claim has three parts. First, with footprints taken from harmonic cycle representatives, the matching agrees with Betti matching's induced pairs on a large majority of features, compared with the under-1% agreement implied by Stucki et al.'s report on Wasserstein matching. Second, training with it reaches Betti-matching-level topological error on full images instead of patches, at a fraction of the wall-clock cost. Third, on point clouds with known ground-truth correspondence, it matches loops correctly under noise where birth–death matching fails.
This is not a new descriptor. It is a cheap change to the matching step, paired with the validation that the enriched-topology literature has mostly skipped.
Relation to Prior Work
Two results motivate the proposal by showing that birth–death summaries lose geometry that matters. Carrière and Oudot show that the bottleneck distance is only a pseudo-metric on Reeb graphs: it can give distance zero to different structures. Stucki et al. show that this blindness corrupts the gradients of a training loss. Adams and Coskunuzer prove that a feature's lifespan is governed by its geometric width. Lifespan therefore says nothing about position, which is exactly the information our footprint term adds. Their results are purely theoretical, and we do not test them directly.
The method borrows two pieces. Footprints come from harmonic representatives, the zero-eigenvectors of the Hodge Laplacian (a discrete analogue of the Laplace operator on edges and triangles). Wee et al. used these representatives to locate molecular loops and compared them through ultrametric Gromov–Hausdorff distances on 900 perovskite structures. We use the same representatives for a different purpose: to place features in space inside a differentiable matching loss, not to compare whole molecules. The exact assignment step follows the lesson from alignment work such as SimAlign and ALOHa: global Hungarian matching beats greedy pairing. Both of those works, however, treat their units as flat sets with no topological structure.
One result constrains the design. Davies et al. show that richer spectral-topological features beat plain persistent homology on MNIST (82.1% vs 62.5%) but trail CNNs (>99%). This is why we do not propose enriched features as model inputs. We change only a loss that sits on top of a standard network.
Finally, ToMCLIP aligns embedding spaces across languages using only connected components ($H_0$). Our point-cloud setting targets loops ($H_1$), which it leaves out.
Research Program
We will build a GPU implementation with four stages. It computes cubical or Vietoris–Rips persistence of the prediction and the target. It extracts, for each feature, a representative cycle: either the harmonic representative, or a cheap stand-in given by the critical cell's coordinates. It computes pairwise footprint distances; entropic optimal transport is enough, because each footprint lives on only a few hundred cells. It then solves the feature-level Hungarian assignment. Gradients flow through birth and death values at the critical cells in the usual way. The matching itself and the footprints are held fixed during backpropagation.
The analysis has two parts. The first is agreement with induced matching on image pairs where Betti matching is defined. Here Betti matching serves as a reference answer, and we measure how agreement changes with $\lambda$, feature dimension and noise. The second is where the proxy fails. We expect it to fail when two nested or adjacent cycles have overlapping footprints, the case where only the algebraic induced map can tell them apart. We will measure how often that configuration occurs on real data rather than assume it is rare. Every experiment reports wall-clock time and memory alongside accuracy, which is the empirical-validation gap this work is meant to close.
Baselines
The naive reference is a U-Net trained with Dice loss alone, plus clDice (a centerline-overlap loss used as a baseline in Stucki et al.). Both are trained with the same architecture, schedule and augmentation. They set the floor that any topological term must beat on topological error without losing Dice.
The spatially blind reference is the standard Wasserstein persistence-diagram loss, which is our own method at $\lambda=0$. It isolates the effect of the location term. It runs in the same code path, so compute is identical.
The strongest competitor is Betti matching with the DiceBetti loss from Stucki et al. (ICML 2023, and the 3D version from 2024). It is the only existing method that matches features by spatial correspondence, and it was shown to beat Dice, clDice and Wasserstein losses on CREMI, road networks, synthetic MNIST, C. elegans and colon-cell images. We run the authors' implementation in two ways: their 48×48 patch protocol, and the largest full-image or volume size it can handle within the same GPU-hour budget as our method. Loss weights for both methods are tuned with the same number of trials on a shared validation split.
There is no reported Betti matching error number on CREMI or the other 2D sets on Lacuna. The only quantitative anchor in the file is that Wasserstein matching mismatches spatially more than 99% of the time. We will therefore reproduce Betti matching's numbers ourselves. We count our method as beating it if it achieves equal or lower Betti matching error $\mu^{err}$ (Betti matching's own metric, which favors the baseline) at matched Dice with at least 3× lower training wall-clock. Alternatively, a strict improvement in $\mu^{err}$ once full-image training becomes affordable also counts.
Two ablations isolate the key ingredient. The first replaces harmonic footprints with critical-cell coordinates, which tests whether a full cycle representative is needed. The second replaces the Hungarian assignment with greedy nearest-footprint matching, which tests whether global optimality matters once location is in the cost.
Why not just…?
Why not just attach the birth and death pixel coordinates to each diagram point and match on those? This is the cheapest version of our idea and a real threat. It is also the ablation we run. Critical cells can jump far under tiny perturbations, because a loop's birth pixel can be anywhere on the loop. So we expect it to be noisier than harmonic footprints, especially for large loops. If the coordinate ablation matches harmonic footprints in agreement with Betti matching and in trained $\mu^{err}$, we will adopt it as the recommended method and report the harmonic version as unnecessary.
Why not just use Betti matching everywhere? On grids it is the right answer when it is affordable. Its cubic pixel cost is what forced 48×48 patches and CPU-bound 3D training, and our budget-matched comparison measures that tradeoff directly. Off grids it needs a shared ambient complex. That complex exists for meshes only when vertex correspondence is already known, and correspondence is what one usually lacks. In the mesh experiment we will build that shared complex from the ground-truth correspondence and run induced matching as an oracle upper bound. Our method does not get to see that correspondence.
Why not just use greedy nearest-location matching instead of Hungarian? Greedy pairing is known to cascade errors when an early unit claims a shared partner (Sultan et al.). SimAlign found that global matching beats argmax alignment. With only tens to hundreds of features per image, the exact assignment costs little, so the greedy ablation should matter only when features crowd together. The experiment will show whether it does.
Why not just rely on Dice or clDice? Stucki et al. report that Dice-trained networks create false bridges and miss one-pixel breaks with little change in Dice. Topology-aware losses fix this without lowering Dice. The Dice-only baseline is included so that this claim is re-measured under our protocol rather than assumed.
Proposed Experiments
The first experiment checks agreement. On CREMI and the four other 2D datasets from Stucki et al., we take predictions from checkpoints at several training stages. For each, we compute Betti matching's induced pairs and our pairs, and report the fraction of induced pairs we recover, split by homology dimension and by $\lambda$. It is decisive if agreement exceeds 90% for harmonic footprints while $\lambda=0$ stays near the reported under-1% spatial correctness. If agreement stays below roughly 70%, geometric proximity is not a usable substitute for induced maps. We would report that negative result and stop the loss experiments.
The second experiment trains segmentation models. We train on the same 2D datasets, then on VesSAP and BBBC027 in 3D. We report $\mu^{err}$, Betti number errors, Dice, and GPU-hours, comparing patch training with full-image or full-volume training. The test is whether whole-image training, which our method makes cheap, lowers topological errors that come from structures spanning patch borders, such as long vessels and neurites.
The third experiment moves off the grid. We sample noisy point clouds from DeformingThings4D-Matching pose pairs, 433 non-isometric mesh pairs with ground-truth vertex correspondences (Magnet et al.). Poses in which limbs touch the body create $H_1$ loops. We compute Vietoris–Rips $H_1$ features and label a pair of loops as corresponding when their representatives map onto each other under the known vertex map. We then measure matching accuracy against noise level for $\lambda=0$, critical-cell footprints, harmonic footprints, greedy matching, and the oracle induced matching. Success means harmonic footprints stay close to the oracle while $\lambda=0$ falls toward chance as the noise grows. No reported result was found on Lacuna for persistence-feature correspondence on this dataset, so all numbers here will be new and reported with seeds and runtime.
References
- Topologically Faithful Image Segmentation via Induced Matching of Persistence Barcodes
- Efficient Betti Matching Enables Topology-Aware 3D Segmentation via Persistent Homology
- Local Equivalence and Intrinsic Metrics between Reeb Graphs
- Geometric Approaches to Persistent Homology
- A cohomology-based Gromov–Hausdorff metric approach for quantifying molecular similarity
- SimAlign: High Quality Word Alignments without Parallel Training Data using Static and Contextualized Embeddings
- ALOHa: A New Measure for Hallucination in Captioning Models
- The Persistent Laplacian for Data Science: Evaluating Higher-Order Persistent Spectral Representations of Data
- Topological Alignment of Shared Vision-Language Embedding Space
- Feature-Rich Two-Stage Logistic Regression for Monolingual Alignment
- Smooth Non-Rigid Shape Matching via Effective Dirichlet Energy Optimization