MATCH: Multi-faceted Adaptive Topo-Consistency for Semi-Supervised Histopathology Segmentation

Conference: NeurIPS 2025
arXiv: 2510.01532
Code: GitHub
Area: Medical Imaging
Keywords: Semi-supervised segmentation, topological consistency, persistent homology, histopathology, MC dropout

TL;DR

This paper proposes the MATCH framework, which tightly couples topological reasoning with the perturbation-robustness principle of semi-supervised learning. By exploiting dual-level topological consistency across random perturbations and temporal training snapshots, MATCH adaptively identifies reliable topological structures without requiring manually defined thresholds, substantially reducing topological errors in histopathology image segmentation.

Background & Motivation

Accurate segmentation of glands and cell nuclei in histopathology images is critical for digital pathology, directly affecting diagnosis, prognosis, and treatment planning. Two major challenges arise:

Topological errors: Densely packed cellular structures frequently cause topological mistakes (e.g., false merges or false splits). Even when pixel-level metrics appear acceptable, topologically incorrect predictions severely undermine clinical reliability.

Annotation cost: Fully supervised methods require large amounts of annotated data, which is prohibitively expensive and unscalable in histopathology, motivating the exploration of semi-supervised learning (SSL) strategies.

Existing SSL methods generally do not explicitly optimize for topological errors. TopoSemiSeg was the first to incorporate topological constraints into an SSL framework, leveraging persistent homology to enforce topological consistency between teacher and student predictions. However, its fundamental limitation lies in relying on predefined, manually chosen persistence thresholds to determine which topological structures are meaningful. Such fixed thresholds are not data-driven and may miss low-persistence yet genuinely meaningful structures, or retain high-persistence noise.

The core insight of this work is that the fundamental principle of SSL is robustness to perturbations—pixel-level predictions that persist across different perturbations are considered reliable. Elevating this principle from the pixel level to the topological level, topological structures that consistently appear across different perturbations are regarded as reliable. This approach eliminates the need for manual thresholds and instead adaptively identifies truly meaningful structures.

Method

Overall Architecture

MATCH adopts a teacher–student framework with the following total loss:

\[\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{sup}} + \lambda_{\text{cons}}\mathcal{L}_{\text{cons}} + \lambda_{\text{intra}}\mathcal{L}_{\text{intra}} + \lambda_{\text{temp}}\mathcal{L}_{\text{temp}}\]

where \(\mathcal{L}_{\text{sup}}\) is the Dice + cross-entropy loss on labeled data, \(\mathcal{L}_{\text{cons}}\) is the pixel-level consistency loss, and \(\mathcal{L}_{\text{intra}}\) and \(\mathcal{L}_{\text{temp}}\) are the proposed dual-level topological consistency losses.

Key Designs

1. MATCH-Pair: Spatially-Aware Pairwise Matching Algorithm: Existing topological matching methods have notable shortcomings—Wasserstein matching relies solely on persistence without considering spatial relationships, while Betti matching incorporates spatial context but still produces errors in the absence of ground truth. MATCH-Pair designs a matching metric combining three factors:

   \(S_{ij} = w_{1,i} \cdot w_{2,j} \cdot \frac{|M_{1,i} \cap M_{2,j}|}{|M_{1,i} \cup M_{2,j}|} \cdot \left(1 - \frac{d_{ij}}{d_{\max}}\right)\)

   • \(w_{k,i}\): normalized persistence weight (reflecting topological importance)
   • IoU term: spatial overlap
   • Distance penalty: Euclidean distance between birth critical points

A flood-fill algorithm expands from birth pixels to generate spatial masks \(M_i\) for each persistence pair. The Hungarian algorithm then solves for the globally optimal one-to-one matching.

1. MATCH-Global: Multi-faceted Global Matching: Extends pairwise matching to global matching across multiple random predictions (facets). MATCH-Pair is applied to each adjacent facet pair \((t, t+1)\); all matchings form an undirected graph \(G = (\mathcal{V}, \mathcal{E})\), and breadth-first search identifies connected components \(\{\mathcal{C}_k\}_{k=1}^K\), each representing a globally consistent topological structure identity.

2. Dual-Level Topological Consistency: Two complementary consistency levels are enforced:

   • Intra-topological consistency: across multiple MC dropout random predictions of the same input, ensuring topological consistency under different model perturbations.
   • Temporal-topological consistency: across consecutive training snapshots, ensuring topological consistency across different training stages.

Matched structures are encouraged toward optimal probability distributions (birth→0, death→1), while unmatched structures are driven toward shorter topological lifetimes (reducing the birth–death gap):

$\mathcal{L}_{\text{match}}(t,i) = (P_{b_{t,i}}^{(t)})^2 + (1 - P_{d_{t,i}}^{(t)})^2$
$\mathcal{L}_{\text{diag}}(t,i) = (P_{b_{t,i}}^{(t)} - P_{d_{t,i}}^{(t)})^2$

Loss & Training

• The teacher model is updated via EMA: \(\theta_t^{(\tau+1)} = \alpha\theta_t^{(\tau)} + (1-\alpha)\theta_s^{(\tau+1)}\)
• The student receives strongly augmented inputs; the teacher receives weakly augmented inputs.
• \(B_{\text{intra}} = B_{\text{temp}} = 4\) is the optimal number of facets.
• Persistent homology uses super-level set filtration to extract 0-D topological features.

Key Experimental Results

Main Results: Three Histopathology Datasets

Dataset	Label Ratio	Method	Dice_Obj↑	BE↓	BME↓	DIU↓
CRAG	10%	TopoSemiSeg	0.884	0.227	10.475	49.690
CRAG	10%	Ours	0.888	0.197	9.175	45.950
CRAG	20%	TopoSemiSeg	0.898	0.226	8.575	43.712
CRAG	20%	Ours	0.909	0.188	7.425	40.250
GlaS	10%	TopoSemiSeg	0.878	0.551	8.300	35.845
GlaS	10%	Ours	0.884	0.501	7.850	30.525
MoNuSeg	20%	TopoSemiSeg	0.793	5.150	188.642	1105.946
MoNuSeg	20%	Ours	0.790	4.930	179.225	982.286

Ablation Study

Configuration	Dice_Obj↑	BE↓	BME↓	DIU↓
w/o \(\mathcal{L}_{\text{intra}}\) & \(\mathcal{L}_{\text{temp}}\)	0.862	0.460	11.680	59.930
\(\mathcal{L}_{\text{intra}}\) only	0.898	0.215	7.920	44.750
\(\mathcal{L}_{\text{temp}}\) only	0.882	0.238	8.540	45.310
\(\mathcal{L}_{\text{intra}}\)+\(\mathcal{L}_{\text{temp}}\)	0.909	0.188	7.425	40.250
Wasserstein matching	0.864	0.423	9.647	58.592
Betti matching	0.889	0.237	8.216	44.157
Ours matching	0.909	0.188	7.425	40.250

Key Findings

• MATCH outperforms TopoSemiSeg on topological metrics across all three datasets while maintaining comparable or superior pixel-level metrics.
• \(\mathcal{L}_{\text{intra}}\) and \(\mathcal{L}_{\text{temp}}\) each independently improve performance, with their combination achieving the best results.
• The MATCH-Pair matching algorithm substantially outperforms Wasserstein and Betti matching (BE: 0.188 vs. 0.423 vs. 0.237).
• Four facets for MC dropout is optimal—too few reduces reliability, while too many introduces redundancy and noise.
• Uncertainty estimation naturally emerges as a by-product of topological consistency, with Pearson correlation between prediction error and uncertainty exceeding 0.72.
• MATCH maintains significant advantages over baselines even in densely packed nuclear regions (≥100 nuclei).

Highlights & Insights

• The core idea is elegant: elevating the SSL perturbation-robustness principle from the pixel level to the topological level is a natural and powerful extension.
• Adaptive identification of topological structures without manual thresholds is more general and data-driven than fixed-threshold approaches.
• The MATCH-Pair metric, integrating spatial overlap, persistence weights, and spatial distance, is well-designed and comprehensive.
• The dual-level consistency (perturbation dimension + temporal dimension) is complementary, covering topological uncertainty from distinct sources.
• The natural emergence of uncertainty estimation constitutes a valuable ancillary benefit.

Limitations & Future Work

• The method primarily focuses on 0-D topological features (connected components); 1-D features (loops/holes) are only preliminarily validated on the Roads dataset.
• The matching algorithm depends on persistent homology computation, which may incur substantial overhead for large-scale, high-resolution whole-slide images.
• The optimal values of \(B_{\text{intra}}\) and \(B_{\text{temp}}\) may vary across datasets and are not determined adaptively.
• The framework could be extended to other medical image segmentation tasks (e.g., vessel segmentation, cell tracking).
• Replacing MC dropout with more efficient uncertainty estimation methods may further improve computational efficiency.

Related Work & Insights

• The application of persistent homology to topology-aware segmentation is increasingly mature; this work elegantly integrates it with SSL principles.
• Relation to TopoSemiSeg: MATCH inherits the overarching direction of topology constraints + SSL, while addressing the critical limitation of fixed thresholds.
• The design of MATCH-Pair can inspire other tasks requiring cross-prediction structure matching (e.g., feature correspondence in multi-view 3D reconstruction).
• The idea of temporal topological consistency can be generalized to self-supervised signal design during model training.

Rating

• Novelty: ⭐⭐⭐⭐⭐ The coupling of topological reasoning with the SSL perturbation-robustness principle is highly innovative; the dual-level consistency and comprehensive matching algorithm represent solid technical contributions.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Three datasets, two labeling ratios, seven baseline methods, and multiple ablation studies provide comprehensive coverage.
• Writing Quality: ⭐⭐⭐⭐ The methodology is presented with clear logic, consistent mathematical notation, and effective visualizations.
• Value: ⭐⭐⭐⭐⭐ The work has direct and significant practical value for topologically accurate segmentation in histopathology under low-annotation settings.

Related Papers

• [AAAI 2026] Graph-Theoretic Consistency for Robust and Topology-Aware Semi-Supervised Histopathology Segmentation
• [NeurIPS 2025] VQ-Seg: Vector-Quantized Token Perturbation for Semi-Supervised Medical Image Segmentation
• [NeurIPS 2025] STARC-9: A Large-scale Dataset for Multi-Class Tissue Classification for CRC Histopathology
• [NeurIPS 2025] Domain-Adaptive Transformer for Data-Efficient Glioma Segmentation in Sub-Saharan MRI
• [NeurIPS 2025] UniMRSeg: Unified Modality-Relax Segmentation via Hierarchical Self-Supervised Compensation