Learning Topology-Aware Representations via Test-Time Adaptation for Anomaly Segmentation

Ali Zia, Usman Ali, Kang Han, Abdul Rehman, Muhammad Umer Ramzan, shahnawaz qureshi, Wei Xiang

Keywords: Topological Data Analysis, Test Time Training, Multi Level Filterisation

Abstract: Test-time adaptation (TTA) has emerged as a powerful paradigm for handling distribution shifts in deep models, particularly for anomaly segmentation, where pixel-wise labels of anomalous regions are typically unavailable during training. We introduce TopoTTA (Topological Test-Time Adaptation), a novel framework that incorporates persistent homology, a tool from topological data analysis, into the TTA pipeline to enforce structural consistency in segmentation. By applying multi-level cubical complex filtrations to anomaly score maps, TopoTTA generates robust topological pseudo-labels that guide a lightweight test-time classifier, enhancing binary segmentation quality without retraining the backbone model. Our method eliminates the need for heuristic thresholding and generalises across both 2D and 3D modalities. Extensive experiments on MVTec AD and BraTS datasets demonstrate significant improvements over state-of-the-art unsupervised anomaly detection and segmentation methods in terms of F1 score, particularly on anomalies with complex geometries.

Primary Area: General machine learning (supervised, unsupervised, online, active, etc.)

Submission Number: 14791