Source-linked AI summary
Asymmetric Paired-Annotation Learning for Multi-Structure ULF Pediatric Brain MRI Segmentation
Ha-Hieu Pham, Dang P. M. Cao, Minh Hoang Pham, Khanh Nguyen Vo Ngoc, Thanh-Huy Nguyen, Ulas Bagci, Huy-Hieu Pham
TL;DR
Pediatric segmentation from portable 0.064 T ULF MRI is difficult because image quality limits anatomical visibility and HF-derived references can be locally misregistered. The paper introduces AURA, which anchors supervision to the scored HF mask while reliability-gating LF information, and reports small development-set differences with preliminary evidence of complementary benefit. Further evaluation is needed because the method was assessed in a challenge development setting with limited separation among strong submissions.
Problem
ULF MRI expands access to pediatric neuroimaging, but segmentation is challenged by weak image quality, non-equivalent HF and LF annotations, and local HF registration errors.
Method
AURA uses HF masks as primary supervision and incorporates LF masks through a bounded reliability gate in an nnU-Net-based paired-supervision strategy with ULF-only inference.
Results
The final ensemble achieved the best development-set DSC and ASSD, but its DSC gain over OS50 was only 0.0004, from 0.7984 to 0.7988.
Takeaways & Limitations
The results provide preliminary evidence that LF-edited annotations can supply complementary image-aligned information when used as bounded auxiliary supervision.
Takeaways & Limitations
The small separation among high-performing validation submissions limits claims of practically meaningful superiority, and hidden-validation results are not directly comparable to the 16-case development results.
Abstract
from arXiv · showhide
Portable ultra-low-field (ULF) MRI can expand access to pediatric neuroimaging, but segmentation at 0.064 T remains challenging because anatomical boundaries are weakly delineated, small structures may be only partially visible, and high-field references can be locally misregistered. The LISA 2026 Challenge provides two non-equivalent annotations reflecting different sources of anatomical evidence: a highfield-derived (HF) mask defining the scored target and a low-field-edited (LF) mask aligned with visible ULF anatomy. In this challenge report, we describe AURA, an nnU-Net-based asymmetric supervision strategy that treats these annotations as distinct observations rather than interchangeable ground truths. AURA anchors training to the HF mask and incorporates the LF mask through a bounded reliability gate based on label disagreement, boundaries, predictive uncertainty, class reliability, and training stage. On a 16-case development split, the HF-supervised baseline, AURA, and their ensemble achieved Dice scores of 0.7984, 0.7950, and 0.7988, respectively, while the ensemble achieved an HD95 of 1.8892 and an ASSD of 0.7855. These results provide a preliminary evaluation of AURA within the LISA 2026 Challenge and motivate further assessment on the hidden test set and external ULF cohorts. Our code and pretrained models are available at https://github.com/minhdang050806/ A-nnU-Net-based-asymmetric-supervision-strategy.
1 Introduction
Portable ULF MRI improves access to pediatric neuroimaging but makes multi-structure segmentation difficult, while paired HF and LF annotations provide non-equivalent anatomical evidence. AURA retains HF supervision as primary and uses LF information through bounded, reliability-aware supervision with ULF-only inference.
- At 0.064 T, lower signal-to-noise ratio, weaker tissue contrast, partial-volume effects, and motion sensitivity complicate pediatric brain segmentation.
- HF masks define the scored target but may contain registration errors, whereas LF masks reflect visible ULF anatomy but may omit or simplify poorly resolved structures.
- AURA uses a bounded reliability gate to incorporate LF edits while retaining the HF mask as the primary supervision source.
- AURA is initialized from the HF-supervised baseline and fine-tuned with asymmetric paired-label supervision.
- The report evaluates a reproducible nnU-Net-compatible training, ensembling, and post-processing recipe on the available development setting, not independent generalization.
2 Challenge Overview
LISA 2026 Task 2 evaluates 11-class pediatric brain segmentation from 0.064 T ULF MRI using HF-derived scored targets and optional LF masks aligned with visible anatomy. The annotations are complementary but non-equivalent, motivating paired-supervision analysis and multi-axis evaluation.
- 2.1 Task Definition and Released Data: LISA 2026 Task 2 uses combined isotropic T2-weighted ULF volumes acquired at 0.064 T and includes 11 foreground labels.
- 2.1 Task Definition and Released Data: The dataset provides an HF-derived official target and an LF-edited annotation aligned with structures visible in the ULF image.
- 2.1 Task Definition and Released Data: The fixed split contains 63 training cases and 16 development-validation cases, with both annotations available for each training case.
- 2.2 Paired-Annotation Structure: The HF and LF masks are related but not equivalent because HF registration can displace labels while LF editing can omit or simplify poorly resolved anatomy.
- 2.2 Paired-Annotation Structure: 38.1% disagreement occurred in the highest-mismatch training case, with differences concentrating near anatomical boundaries and sometimes affecting most of a structure on a slice.
- 2.1 Task Definition and Released Data: Evaluation uses DSC, HD, HD95, ASSD, signed RVE, and absolute RVE to cover overlap, boundary error, surface distance, and volume bias.
3 Methodology
AURA uses HF supervision as the anchor and incorporates LF annotations through bounded, reliability-aware asymmetric supervision within an unchanged nnU-Net framework. Independently trained OS50 and AURA predictors are fused at inference using probability averaging and connected-component filtering.
- 3.1 Paired-annotation training: OS50 is trained directly on HF-derived targets, while AURA is initialized from OS50 and fine-tuned with paired HF/LF supervision.
- 3.2 Asymmetric supervision: LF influence is gated using HF/LF disagreement, label-boundary membership, predictive entropy, class-specific reliability, and training stage.The epoch ramp delays LF supervision until after a 25-epoch HF-only warm-up.
- 3.2 Asymmetric supervision: The multiplicative gate keeps LF contribution bounded by 0.42, ensuring the HF anchor retains the majority of target mass.Detached reliability signals prevent the network from changing the gate through gradient flow.
- 3.2 Asymmetric supervision: AURA treats LF labels as a conditional auxiliary observation rather than a second ground truth.The HF-derived target remains the primary supervision source.
- 3.2 Asymmetric supervision: AURA combines class-weighted soft Dice, soft TopK10 cross-entropy, and a boundary-restricted soft Dice term.The boundary term is weighted by 0.03 in the reported objective.
- 3.3 Probability fusion and post-processing: At inference, OS50 and AURA receive only ULF MRI, their probability maps are averaged, and foreground predictions undergo largest-connected-component filtering.The 0.6/0.4 ensemble weight was selected on the development split and is validation-tuned.
4 Experiments and Results
The experiments evaluate reproducible nnU-Net training and compare OS50, AURA, and ensemble predictions on 16 development cases. The ensemble provides only modest overlap and surface-distance changes over OS50, while qualitative gains concentrate near boundaries and small peripheral regions.
- Quantitative Comparison: The development evaluation uses macro-averaged results over 16 cases and 11 foreground labels.Reported metrics cover overlap, boundary error, surface distance, and volume bias.
- Quantitative Comparison: 0.7988 ensemble DSC is the best reported DSC, compared with 0.7984 for OS50 and 0.7950 for AURA.The ensemble also achieves ASSD 0.7855, while HD95 does not improve over OS50.
- Qualitative Analysis: +0.00326 is the largest positive DSC change over OS50 among the illustrated cases.The improved case shows refinements confined to boundary regions where the baseline already recovers the overall anatomy.
- Qualitative Analysis: The lowest-DSC case shows broader errors across bilateral structures, with fusion changing performance only marginally.Its ensemble improvement is DSC +0.0001 over OS50, and AURA does not address this failure mode.
5 Limitations
The limitation analysis places AURA’s results within the LISA 2026 challenge setting rather than treating them as evidence of broad superiority. Leaderboard separation is narrow, development and hidden-validation results are not directly comparable, and further evaluation is required.
- Challenge Scope: Many leading submissions fall within a narrow DSCAvg range of approximately 0.82–0.83.The OS50+AURA submission has DSCAvg = 0.82 and HDAvg = 3.41, near the leading end of both distributions.
- Challenge Scope: The hidden-validation leaderboard values are not directly comparable to the 16-case development results in Table 1.They are computed on hidden validation cases under the challenge setting.
- Scope Boundaries: The evaluation does not establish practically meaningful superiority, robustness, or generalization beyond the challenge setting.The current analysis also does not isolate individual reliability-gate factors.
6 Conclusion
The conclusion presents AURA as an asymmetric supervision strategy for 11-class ULF pediatric brain MRI segmentation. It uses HF annotations as the primary target and LF annotations as bounded auxiliary supervision during training, but its evidence remains preliminary and challenge-limited.
- Conclusion: AURA is an asymmetric dual-observation strategy for 11-class ULF pediatric brain MRI segmentation.The method treats HF-derived and LF-edited annotations as distinct observations rather than interchangeable ground truths.
- Conclusion: The model operates solely on ULF MRI at inference without requiring annotations or a high-field image.Paired annotations are used only during training.
- Conclusion: The development results provide preliminary evidence that paired annotations may offer complementary LF supervision for ULF segmentation.Differences relative to the HF-supervised baseline were small.
- Conclusion: Further evaluation on the hidden test set and external ULF cohorts, with component-wise reliability-gate analysis, is required.The current evaluation does not establish superiority, robustness, or generalization beyond the challenge setting.