PaperlayerView exact revision on arXiv

Source-linked AI summary

Skeleton-Guided Progressive Test-Time Adaptation for Thin Curvilinear Structures

Boa Jang, JunGyu Lee, Gwanho Lee, Jinwook Choi, Young-Gon Kim

arXiv:2610.11104v1cs.CVeess.IV

TL;DR

Thin curvilinear segmentation is vulnerable to broken connectivity under domain shift, especially across imaging modalities. SGP-TTA separately schedules batch-normalization statistics and uses a multi-view skeleton-consensus objective for affine parameters; it achieves strong connectivity results, with its largest gains under cross-modality shift.

  • Problem

    Thin structures are vulnerable to connectivity-breaking pixel errors, while existing test-time adaptation does not constrain connectivity and faces a stage-dependent source–target statistics balance.

    Source: Introduction · p. 1

  • Method

    SGP-TTA schedules the target-statistics weight by samples seen and trains batch-normalization affine parameters with a consensus-skeleton objective from aligned multi-view predictions.

    Source: Introduction · p. 2

  • Results

    SGP-TTA achieves the highest clDice in all 12 transfer settings and the highest Dice in 10 of 12, with its largest average gains under cross-modality shift.

    Source: Key Observation 1: Consistent… · p. 6, Key Observation 2: Larger gains under… · p. 6, Conclusion · p. 7

  • Takeaways & Limitations

    The authors conclude that separating adaptation of the two normalization components matters most under cross-modality transfer, and the useful source–target statistics mixture changes as images accumulate.

    Source: Conclusion · p. 7

    The setting is binary segmentation on an unlabeled target stream; source data and target annotations are inaccessible.

    Source: Problem Setup · p. 3

Abstract

from arXiv · show

Accurate segmentation of thin curvilinear structures is vital for various real-world applications, from vessel analysis to road extraction. Yet their intricate geometry makes even minor pixel-wise errors enough to break the global topology, and this structural fragility turns severe domain shifts into catastrophic failures. The difficulty is most acute under cross-modality gaps, where the imaging process itself differs fundamentally between source and target. While test-time adaptation (TTA) offers a practical source-free remedy, existing methods adapt feature statistics and confidence, neither of which constrains connectivity, and thus degrade under such extreme gaps. To address this, we propose Skeleton-Guided Progressive Test-Time Adaptation (SGP-TTA). Progressive Batch Normalization (ProgBN) shifts normalization from frozen source statistics toward current target estimates under a sample-count schedule, so that the source-target balance follows the stage of adaptation rather than a fixed coefficient. Consensus Skeleton Recall (CSR) then derives a structural target from geometrically aligned multi-view predictions and updates only the BN affine parameters to preserve connected structures. Extensive experiments show that SGP-TTA consistently outperforms existing TTA methods in topological connectivity, with the largest margins under cross-modality shift. The project page is available at https://boa-jang.github.io/SGP-TTA.

Introduction

Thin curvilinear structures are vulnerable to broken connectivity from small pixel errors, and severe domain shifts make segmentation especially challenging. SGP-TTA addresses limits in test-time adaptation by scheduling normalization-statistics updates and supplying structural supervision for affine-parameter updates.

Introduction

  • Small pixel-wise errors can break connectivity in thin, branched structures, undermining downstream applications such as autonomous navigation and biomarker quantification.The challenge spans road extraction and medical vessel segmentation.Source: Introduction · p. 1
  • Existing test-time adaptation lacks a suitable source–target statistics balance across adaptation stages and can suppress weak branches without constraining connectivity.Immediate replacement of source statistics may destabilize early predictions, while a source-heavy mixture can become conservative after affine parameters adapt.Source: Introduction · pp. 1–2
  • SGP-TTA separately adapts batch-normalization statistics and affine parameters through Progressive Batch Normalization and Consensus Skeleton Recall.Statistics are recalibrated from target inputs, while affine parameters are updated through unsupervised optimization.Source: Introduction · p. 2
  • ProgBN monotonically increases the target-statistics weight with the number of observed samples, anchoring early adaptation to source statistics and relying more on target estimates later.Its deterministic schedule needs neither target labels nor offline coefficient learning.Source: Introduction · p. 2
  • CSR aligns and aggregates geometrically transformed predictions, extracts a dilated skeleton from their consensus, and encourages each view to recover its centerline.This self-supervised objective supplies structural supervision absent from confidence-based objectives.Source: Introduction · p. 2

Related Work

Prior work addresses curvilinear geometry and domain shifts through structural supervision, domain adaptation, and test-time updates. The reviewed approaches leave challenges for unlabeled, source-free adaptation that preserves connectivity under severe shifts.

Curvilinear Structure Segmentation

  • Curvilinear segmentation methods encode geometry in architectures or topology-aware losses, while foundation-model adapters extend domain-general segmentation but require reference masks or structural annotations during training.Examples include connectivity and point-set representations, clDice, and skeleton recall loss.Source: Curvilinear Structure Segmentation · p. 2

Cross-Modality Domain Adaptation

  • Cross-modality adaptation uses feature alignment, style–structure disentanglement, or target-style synthesis, while topology-aware road adaptation still depends on source data and offline adversarial alignment.The reviewed approaches generally require pre-collected target images and computationally heavy offline retraining.Source: Cross-Modality Domain Adaptation · p. 2

Test-Time Adaptation

  • Test-time adaptation uses entropy minimization, self-training, or auxiliary self-supervision, but foreground–background imbalance can reinforce errors around weak branches.Aggregating augmented views reduces variability without itself preserving connectivity.Source: Test-Time Adaptation · pp. 2–3
  • Normalization-based TTA replaces or interpolates source statistics, while ProgBN uses a monotone sample-count schedule to blend online target estimates with frozen source statistics.The passage contrasts ProgBN with methods that use fixed or online-estimated interpolation coefficients.Source: Test-Time Adaptation · p. 3

Method

SGP-TTA adapts source-trained binary segmentation models to an unlabeled target stream by treating BN statistics and affine parameters separately. ProgBN recalibrates statistics as adaptation proceeds, while CSR supplies a structural target for affine-parameter updates.

Preliminaries

  • The setup adapts a source-trained binary curvilinear segmentation model to an unlabeled target stream, where source data and target annotations are inaccessible.Under domain shift, BN statistics can be re-estimated from target activations, but affine parameters require an unsupervised objective.Source: Problem Setup · p. 3, Batch Normalization · p. 3

Progressive Batch Normalization (ProgBN)

  • ProgBN progressively increases reliance on current-image target statistics as more target samples are processed, while retaining a positive source-statistics contribution at every finite sample count.The sample count is a label-free proxy for adaptation stage, not pooled statistical precision: each image’s estimated statistics are discarded after processing.Source: Progressive Batch Normalization (ProgBN) · p. 4

Consensus Skeleton Recall (CSR)

  • CSR aligns predictions from geometrically transformed views, extracts a dilated skeleton from their consensus, and encourages each view prediction to recover its supported centerline.The skeleton is derived from the model’s own consensus rather than ground truth, and is treated as a stop-gradient target.Source: Consensus Skeleton Recall (CSR) · p. 4

Overall Objective and Online Adaptation

  • SGP-TTA updates BN affine parameters with one gradient step on a combined CSR and pixel-wise entropy objective, then recomputes the consensus for prediction.The method separates statistic recalibration from affine-parameter optimization; non-BN weights and source statistics remain frozen.Source: Overview · p. 3, Overall Objective and Online Adaptation · p. 4

Experiments

The experiments compare test-time adaptation across retinal and road shifts, evaluating pixel accuracy and topology. SGP-TTA achieves the highest clDice in every transfer setting, with its largest average gains under cross-modality shift.

Experiment Setup

  • Evaluation covers 10 curvilinear segmentation datasets across four imaging domains and three transfer types, reporting Dice and clDice for pixel accuracy and topological continuity.The datasets include retinal CFP, OCTA and FA images, plus road extraction; original dataset splits are used.Source: Datasets and Evaluation Metrics · p. 5, Comparison Results · p. 6
  • All TTA methods adapt from the same source checkpoint in a single-image stream, with each image seen once; UNet is the primary backbone and inputs are resized to 512×512.The baselines include the source-only model and eight TTA methods; SGP-TTA uses six grid-preserving views, one Adam step per target image, and different ProgBN τ values for retinal and road scenarios.Source: Implementation Details · p. 5, Baselines · p. 5

Comparison Results

  • SGP-TTA achieves the highest clDice in all 12 transfer settings and the highest Dice in 10 of 12.On CNDS, it retains 85.27 Dice versus the source model’s 85.56 while raising clDice from 93.39 to 94.30; on MR, it scores 40.80 Dice versus the best 42.24, with the highest clDice at 54.84.Source: Key Observation 1: Consistent… · p. 6
  • Across all four cross-modality transfers, SGP-TTA improves both Dice and clDice over the source; on DRIVE → OCTA3mm it reaches 44.83 Dice and 46.41 clDice, versus best-baseline scores of 32.50 and 34.01.Figure 3 highlights restored fine-vessel continuity and fewer false negatives in the displayed target domains.Source: Key Observation 2: Larger gains under… · p. 6, Figure 3 · p. 6

Analysis and Ablation Studies

  • ProgBN achieves the highest clDice in all three transfer groups and the highest Dice in both retinal groups; the square-root α schedule gives the largest improvement at every point.Figure 4 shows fixed source-heavy normalization leading early and target-only normalization overtaking later, while the square-root schedule performs best throughout.Source: Table 4 · p. 6, Analysis of BN Adaptation Strategies · p. 7, Choice of the α Schedule · p. 7
  • ProgBN provides the largest individual gain among the ablated components, while adding MVC further helps both retinal groups but not uniformly the road setting.Entropy minimization without CSR reduces cross-modality Dice from 47.84 to 35.24; CSR improves both metrics across all three groups.Source: Table 5 · p. 6, Ablation of Adaptation Components · p. 7
  • Across five BN-based backbones, SGP-TTA has the highest mean clDice without architecture-specific modification.Figure 5 averages clDice over all transfer settings for UNet, UNet++, MaNet, LinkNet, and FPN.Source: Robustness Across Segmentation… · p. 7, Figure 5 · p. 7
  • SGP-TTA is slower than entropy-driven baselines but remains below the most expensive compared method while attaining the highest clDice.Figure 6 compares mean inference time and clDice; marker area represents the number of updated parameters.Source: Adaptation Cost · p. 7, Figure 6 · p. 7

Conclusion

SGP-TTA adapts BN statistics and affine parameters through separate mechanisms for thin curvilinear structures in single-image streams. The conclusion reports its greatest importance under cross-modality transfer.

Conclusion

  • Under cross-modality transfer, existing methods degrade below the source model while SGP-TTA improves both metrics.The target modality was never seen during training.Source: Conclusion · p. 7
  • SGP-TTA recalibrates normalization statistics without gradients using an image-count schedule and trains affine parameters against a multi-view skeleton consensus.Source: Conclusion · p. 7
  • The source–target statistics mixture should change as images accumulate, while confidence is a poor proxy for connectivity.The conclusion argues that skeleton-anchored supervision is needed.Source: Conclusion · p. 7
Loading 2610.11104v1…