Source-linked AI summary
CIRSeg: Coarse-to-Fine Intensity-Robust Liver Segmentation with Source-Free Continual Test-Time Adaptation
Ruoshi Xu, Mingqi Gao, Shengda Luo, Jingkun Chen
TL;DR
Limited annotations and heterogeneous MRI intensities make liver segmentation across acquisition domains challenging. CIRSeg combines intensity augmentation, cascaded segmentation, and source-free test-time adaptation; it reports strong CARE 2026 performance and robust generalization to unseen domains.
Problem
Limited annotated volumes and substantial MRI intensity variation make reliable liver segmentation and generalization to unseen acquisition domains challenging.
Source: §1 · p. 2Method
CIRSeg combines 3D CutMix and stochastic Nyul or histogram-matching augmentation with cascaded nnU-Netv2, source-free continual test-time adaptation, and largest-component filtering.
Source: §1 · p. 2, §2.2 · p. 4, §2.3 · p. 5Results
CIRSeg achieves strong performance on the CARE 2026 test set, demonstrating accurate segmentation and robust generalization to unseen acquisition domains.
Source: §1 · p. 3, §4 · p. 10Takeaways & Limitations
The reported results support accurate liver segmentation across in-domain and unseen acquisition settings.
Source: §1 · p. 3No paired no-TTA test predictions were available for causal comparison, and test-order sensitivity of continual TTA could not be evaluated.
Source: §3.4 · p. 10
Abstract
from arXiv · showhide
Reliable liver segmentation in contrast-enhanced MRI is essential for quantitative hepatic assessment, treatment planning, and longitudinal disease monitoring. However, limited annotated data and scanner- or vendor-dependent intensity variations can cause overfitting and poor generalization to unseen acquisition domains. Moreover, simultaneously achieving robust global localization and precise boundary delineation remains challenging, while predictions may contain isolated false-positive regions outside the main liver component. To address these challenges, we propose CIRSeg, a coarse-to-fine, intensity-robust liver segmentation framework based on nnU-Netv2. CIRSeg combines 3D CutMix with stochastic intensity transfer using either Nyul augmentation or histogram matching to improve robustness to heterogeneous MRI intensities. Its cascaded architecture decouples low-resolution anatomical localization from full-resolution boundary refinement. At inference, source-free test-time adaptation based on confidence-filtered predictions and probability-prior regularization further improves robustness to out-of-distribution inputs. As a final deterministic post-processing step, largest connected component filtering removes isolated false-positive regions. On the CARE 2026 test set, CIRSeg achieves Dice scores of 97.13\% and 97.93\% on the in-domain and unseen-domain subsets, with corresponding HD95 values of 20.18 mm and 11.30 mm, respectively. These results demonstrate consistently accurate segmentation across both in-domain and unseen acquisition settings. The code is available at https://github.com/jingkunchen/MICCAI_CARE_2026
1 Introduction
CIRSeg addresses limited annotations and heterogeneous MRI appearances with intensity-robust augmentation, coarse-to-fine segmentation, and source-free test-time adaptation. Its introduction frames these components as coordinated responses to challenges often studied separately.
1 Introduction
- CIRSeg targets limited training diversity, scanner-dependent appearance variation, the need for global localization and precise boundaries, and disconnected false positives.Source: §1 · p. 2
- The introduction argues that annotation scarcity, intensity variation, multi-scale modeling, and inference-time shift require coordinated treatment, although related approaches often address them independently.Source: §1 · p. 2
- CIRSeg combines 3D CutMix with probabilistic Nyul augmentation or histogram matching to increase spatial and intensity diversity under limited annotations.Histogram-matching references are randomly sampled from the training set.Source: §1 · pp. 2–3
- A cascaded nnU-Netv2 separates low-resolution liver localization from full-resolution refinement, while source-free continual adaptation uses confidence-filtered predictions and probability-prior regularization.The adaptation is intended to improve robustness to out-of-distribution inputs while limiting drift.Source: §1 · pp. 2–3
- A deterministic largest connected component filter removes isolated false-positive regions and retains the dominant liver component.Source: §1 · p. 3
2 Method
CIRSeg builds on CoSSeg-TTA and combines composite data augmentation, cascaded nnU-Netv2, and source-free continual test-time adaptation. Its pipeline separates coarse liver localization from fine boundary refinement and applies deterministic largest-component filtering.
2.1 Baseline
- The baseline retains CoSSeg-TTA’s nnU-Netv2, mean-teacher semi-supervised learning, and histogram matching components, then focuses on data diversity, coarse-to-fine segmentation, and inference-time robustness.CoSSeg-TTA was the first-place solution of the CARE Liver 2025 challenge.Source: §2.1 · p. 3
2.2 Composite Data Augmentation
- 3D CutMix applies the same sampled cuboid to two labeled volumes and their labels, preserving voxel-wise supervision while increasing local appearance and boundary diversity.It is used during supervised training to improve regularization under limited annotations.Source: §2.2 · p. 4
- Intensity augmentation probabilistically applies histogram matching or Nyul augmentation using a randomly sampled training-set reference, including unlabeled cases, while leaving labels unchanged.Nyul augmentation blends mapped intensities with the source using transfer strength α; histogram matching transfers the source distribution toward the reference.Source: §2.2 · p. 4
2.3 Cascaded Training
- The cascade first predicts coarse liver extent from a resampled low-resolution volume, then supplies that prediction as an anatomical prior to the full-resolution model for local structure and boundary refinement.The coarse prediction is resampled to the original image space before being provided to the full-resolution cascade.Source: §2.3 · p. 5
2.4 Inference-Time Adaptation and Refinement
- At inference, a frozen anchor and adaptive student support source-free continual test-time adaptation of the full-resolution cascade using confidence-filtered anchor predictions and probability-prior regularization.The student’s adapted parameters carry forward across test volumes, while the anchor remains frozen.Source: §2.4 · p. 6
- Deterministic 3D largest connected component filtering leaves predictions with at most one foreground component unchanged and removes isolated false-positive regions otherwise.The method is described as improving anatomical plausibility, particularly for surface-distance metrics such as HD95.Source: §2.4 · pp. 6–7
3 Experiments
Experiments compare CIRSeg with other segmentation methods and assess its training, architecture, and adaptation components. Ablations report improvements in HD95 from connectivity filtering, augmentation, and cascading, while test-time adaptation shows domain-dependent utility.
3.1 Dataset and Evaluation Metrics
- The official evaluation uses the provided training, validation, and test sets, with Dice and HD95 as its metrics.Higher Dice and lower HD95 indicate better performance.Source: §3.1 · p. 7
3.2 Implementation Setup
- Both nnU-Netv2 stages use a mean-teacher objective, while augmentation combines 3D CutMix with probabilistic histogram matching or Nyul augmentation.The full-resolution cascade stage enables source-free continual test-time adaptation; the low-resolution stage does not.Source: §3.2 · pp. 7–8
3.3 Main Experiment
- CIRSeg achieves the highest Dice score and lowest HD95 among evaluated methods in five-fold cross-validation on 30 annotated training cases.The comparison uses no pretrained models; CIRSeg and CoSSeg-TTA also use unlabeled cases through semi-supervised components.Source: §3.3 · p. 8, Table 1 · p. 9
- CIRSeg predictions show more complete, anatomically consistent liver regions, smoother boundaries, and fewer disconnected false positives than previous approaches.The qualitative comparison uses representative validation cases because annotations are unavailable.Source: §3.3 · p. 8, Figure 3 · p. 9
3.4 Ablation Study
- HD95 falls from 24.69 to 20.61 with largest connected component filtering, to 19.18 after 3D CutMix, and improves further with composite intensity augmentation.The baseline has strong Dice performance but relatively high HD95, indicating boundary errors and isolated false positives remain important failure modes.Source: §3.4 · p. 9, Table 2 · p. 10
- The cascade further reduces HD95 to 18.31, while test-time adaptation is lower on ID Test1/2 but higher on OOD.All test submissions used TTA, so no paired no-TTA predictions were available for causal comparison, and ordering sensitivity could not be evaluated.Source: §3.4 · pp. 9–10, Table 2 · p. 10
4 Conclusion
CIRSeg combines intensity augmentation, cascaded nnU-Netv2, source-free continual test-time adaptation, and connected-component filtering for liver segmentation. Experiments on CARE 2026 show accurate segmentation on in-domain and unseen-domain test subsets.
4 Conclusion
- CIRSeg combines 3D CutMix, histogram matching, and Nyul augmentation with cascaded nnU-Netv2, source-free continual test-time adaptation, and largest connected component filtering.The framework targets limited annotations and heterogeneous MRI appearances, separating coarse localization from boundary refinement and removing isolated false positives.Source: §4 · p. 10