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

Abstract:

In the realm of semi-supervised segmentation, extracting significant semantic structures from unlabeled datasets is a critical requirement. This task becomes especially arduous in histopathology image analysis due to the high density of objects within the images. To tackle this difficulty, we present a semi-supervised segmentation framework specifically engineered to reliably identify and maintain pertinent topological features. Our approach utilizes multiple perturbed predictions generated via stochastic dropouts and temporal training snapshots, thereby enforcing topological consistency across these diverse outputs. This consistency mechanism serves to differentiate biologically relevant structures from fleeting and noisy artifacts. A primary hurdle in this procedure is the precise matching of corresponding topological features across predictions when ground truth labels are unavailable. We address this challenge by introducing an innovative matching strategy that combines spatial overlap with global structural alignment, effectively minimizing discrepancies among the predictions. Comprehensive experiments confirm that our method significantly reduces topological errors, yielding more robust and accurate segmentations that are crucial for dependable downstream analysis. The code for this project can be accessed at https://github.com/Melon-Xu/MATCH.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC