PaperlayerView exact revision on arXiv

Source-linked AI summary

Pre- and Post-Treatment Brain Metastases Segmentation Using nnU-Net with Post-Processing for BraTS 2026

Haobin Liu, Xin Wang

arXiv:2609.11477v1cs.CV

TL;DR

The paper addresses volumetric segmentation of pre- and post-treatment brain metastases under heterogeneous imaging, rare resection cavities, and lesion-wise evaluation. It uses a 5-fold nnU-Net ResEnc-L ensemble with LW-DSC-oriented post-processing, finding two robust stages through out-of-fold auditing while a third fails to reproduce beyond the leaderboard.

  • Problem

    BraTS 2026 brain-metastasis segmentation combines substantial inter-site heterogeneity, severe resection-cavity rarity, and a lesion-wise Dice metric dominated by small-lesion detection.

  • Method

    The pipeline combines a 5-fold nnU-Net ResEnc-L ensemble with a rule-based cascade designed for LW-DSC, audited using five-fold out-of-fold evaluation without model-training leakage.

  • Results

    The final pipeline reaches LW-DSC values of 0.733, 0.751, 0.713, and 0.549 for ET, TC, WT, and RC, while RuleClean and BoundaryExpand are robust across folds and brainF70 does not replicate out-of-fold.

  • Takeaways & Limitations

    Under LW-DSC, recall-recovering post-processing carries lower risk than component deletion, and several tested loss, backbone, and inference alternatives degraded performance.

  • Takeaways & Limitations

    Resection-cavity performance remains the sole critical deficit relative to the paper's tumour-side results.

Abstract

from arXiv · show

Brain metastases exhibit high inter-lesion variability in size, enhancement pattern, and post-treatment appearance, making volumetric segmentation of both pre- and post-treatment cases the central challenge of the BraTS 2026 Task 1 (Brain Metastases). We build a pragmatic pipeline on a 5-fold nnU-Net ResEnc-L ensemble, in which each fold is trained independently for 1,000 epochs with the standard Dice + cross-entropy loss on 1,296 four-modality training cases. This ensemble is followed by a rule-based post-processing cascade tuned for the lesion-wise Dice similarity coefficient (LW-DSC), a detection-oriented metric that behaves very differently from the traditional global Dice. The final pipeline reaches an LW-DSC of 0.733 / 0.751 / 0.713 / 0.549 on the enhancing tumour (ET), tumour core (TC), whole tumour (WT), and resection cavity (RC) sub-regions on the official validation leaderboard. Rather than trusting these leaderboard gains, we audit every post-processing stage with a five-fold out-of-fold (OOF) analysis with no model-training leakage over all 1,296 training cases, scored with the official BraTS evaluation code (BraTS_evaluation): it confirms two stages as robust, per-fold-consistent improvements while the third improves only the leaderboard and does not reproduce out-of-fold. We further provide a mechanistic analysis of the LW-DSC metric that explains why recall-recovering post-processing carries low risk whereas component deletion does not, and we report thirteen negative results spanning loss engineering, alternative backbones, and inference-time settings, several of which run counter to widely held intuitions. Source code is released under Apache-2.0 at https://github.com/hornbeamliu/brats2026-met.

1 College of Software, Jilin University, Changchun, China

The authors are affiliated with the Key Laboratory of Symbolic Computation and Knowledge Engineering at Jilin University in Changchun, China.

  • The work is affiliated with Jilin University’s Key Laboratory of Symbolic Computation and Knowledge Engineering in Changchun, China.
  • Keywords identify brain metastases segmentation, nnU-Net, post-processing, lesion-wise Dice, and BraTS 2026.

1 Introduction

The paper addresses volumetric segmentation of pre- and post-treatment brain metastases under heterogeneous data, rare resection-cavity labels, and a lesion-wise evaluation metric. It uses an nnU-Net ensemble with separately auditable post-processing and validates robustness through leakage-free out-of-fold analysis.

  • Brain metastases segmentation supports radiosurgery planning and treatment-response monitoring, but small lesions can be under-detected by manual RANO-BM measurements.
  • BraTS 2026 combines multi-institutional heterogeneity, severe resection-cavity class imbalance, and lesion-wise Dice dominated by detection of small lesions.
  • The pipeline combines a 5-fold nnU-Net ResEnc-L ensemble with separately measurable RuleClean, BoundaryExpand, and brainF70 post-processing stages.
  • A five-fold out-of-fold audit confirms RuleClean and BoundaryExpand as robust, while brainF70 does not reproduce beyond leaderboard behavior.

2 Methods

The dataset contains four MRI modalities and five-class annotations for pre- and post-treatment metastases from eight institutions. Validation uses complete-modality data, while the method adopts a planner-configured 3D full-resolution ResEnc-L nnU-Net.

  • The dataset comprises pre- and post-treatment multiparametric MRI from eight institutions, with four modalities and five-class annotations.
  • The evaluation regions define TC as NETC union ET and WT as NETC union SNFH union ET.
  • All 1,296 training cases provide native T2w volumes, so the pipeline performs no T2w synthesis or zero-imputation.
  • The official 179-case validation set is complete-modality, and local OOF ablations use held-out-fold predictions without validation-submission leakage.
  • The model uses planner-configured 3D full-resolution nnU-Net ResEnc-L with automatically selected network hyperparameters.

2.3 Training

Training uses independently trained ResEnc-L folds with the standard nnU-Net compound loss and default augmentation, followed by separately executable post-processing stages whose effects can be quantified.

  • Each of five folds is trained independently for 1,000 epochs with standard Dice plus cross-entropy loss and default nnU-Net augmentation.
  • Training uses two NVIDIA RTX A6000 GPUs with the planner-selected batch size of 2 for 3D full resolution.
  • The three post-processing scripts run sequentially on the ensemble prediction, allowing each stage’s LW-DSC effect to be quantified separately.

Stage 1: rule-based clean-up (RuleClean).

RuleClean removes small or poorly connected components from the argmax segmentation using four ordered, component-level rules.

  • Stage 1: rule-based clean-up (RuleClean).: RuleClean applies four ordered rules to 26-connected components, relabelling every removed voxel as background.The rules target small ET and RC components, SNFH components lacking tumour-core adjacency, and NETC components lacking ET adjacency.

Stage 2: resection-cavity boundary expansion (BoundaryExpand).

BoundaryExpand grows predicted resection-cavity components conservatively to address systematic boundary under-segmentation while protecting other tumour regions. Its controlled expansion yields a small, consistent RC improvement without changing ET, TC, or WT.

  • Stage 2: resection-cavity boundary expansion (BoundaryExpand).: BoundaryExpand grows each predicted RC component outward in the native voxel grid using background-restricted 3D ring dilation.The stage targets contours truncated inside the true cavity; raw RC lesion-wise DSC is 0.367.
  • Stage 2: resection-cavity boundary expansion (BoundaryExpand).: Algorithm 1 defines the post-cleanup labels and component operations used before BoundaryExpand, including RC as label 4 and the tumour core as labels 1, 3, and 4.The algorithm requires an argmax segmentation and voxel spacing in millimetres.
  • Stage 2: resection-cavity boundary expansion (BoundaryExpand).: The expansion admits background voxels only when ensembled RC probability exceeds 0.10, never enters ET, NETC, or SNFH, and caps each component at 3× its original volume.These three constraints make the operator conservative rather than a blind dilation.
  • Stage 2: resection-cavity boundary expansion (BoundaryExpand).: A small, sign-consistent RC gain appears on every out-of-fold fold, while ET, TC, and WT remain unchanged.This supports BoundaryExpand as a safe, modest operator.

Stage 3: outside-brain false-positive suppression (brainF70).

brainF70 removes resection-cavity components that lie mostly outside the brain using a skull-stripped T1c-derived brain mask, but it fails the out-of-fold audit.

  • Stage 3: outside-brain false-positive suppression (brainF70).: brainF70 deletes any RC component with a brain fraction below 0.70, using B = {v : T1c(v) > 0} without a separate brain-extraction model.Although included in the submitted container, it is excluded from the recommended RuleClean followed by BoundaryExpand cascade because its leaderboard gain does not survive five-fold out-of-fold analysis.

2.5 Inference and ensembling

Inference uses nnU-Net’s standard Gaussian-weighted sliding-window prediction with 50% overlap, and a deterministic fallback handles missing native T2w only during hidden-test inference.

  • 2.5 Inference and ensembling: The predictor uses Gaussian importance weighting with a default step size of 0.5, while a 0.25-step sweep with three ensemble weightings was systematically worse.The five fold softmax volumes are fused by arithmetic mean and decoded with argmax; test-time mirroring across all three axes is enabled.
  • 2.5 Inference and ensembling: Cases lacking native T2w reuse T2-FLAIR for the _0003 channel instead of zero-filling, but this fallback is never exercised because all reported cases provide native T2w.It is retained as a deterministic safeguard for the hidden test phase.

3 Results

The final 5-fold ensemble and post-processing pipeline was evaluated on the official validation set and audited with leakage-free five-fold OOF predictions. The audit identifies RuleClean and BoundaryExpand as robust improvements, while brainF70 does not reproduce reliably.

  • 3.2 Ablation of the post-processing cascade: The OOF audit scores each of the 1,296 training cases once using its held-out fold’s model, avoiding model-training leakage and separating durable effects from leaderboard noise.OOF absolute values are interpreted by stage increment sign and per-fold consistency, not compared directly with ensemble leaderboard values.
  • 3.2 Ablation of the post-processing cascade: RuleClean raises OOF avg4 by +0.0640 and RC by +0.1244, with positive avg4 increments on all five folds.The fold-wise avg4 increment ranges from +0.052 to +0.082.
  • 3.2 Ablation of the post-processing cascade: BoundaryExpand adds +0.0080 RC and +0.0021 avg4 in OOF, remains non-negative across folds, and leaves ET, TC, and WT unchanged.Its RC HD95 also improves from 112.3 to 109.5 mm.
  • 3.2 Ablation of the post-processing cascade: brainF70 changes OOF avg4 by −0.0006 and RC by −0.0020, triggers on only 6 of 1,296 cases, and reverses sign across folds.It is retained in the submitted container only to reproduce the leaderboard score, but excluded from the recommended cascade.
  • 3.2 Ablation of the post-processing cascade: RuleClean improves NSD, HD95, and large-instance F1 across regions but lowers small-instance F1, while BoundaryExpand improves RC HD95 alongside RC overlap.The small-instance F1 decrease is the intended consequence of deleting sub-10 mm3 components.

4 Failed attempts

The failed-attempt analysis finds that imbalance-focused losses, alternative backbones, aggressive component filters, threshold changes, fusion variants, and finer sliding-window overlap did not improve the pipeline reliably. Several approaches substantially harmed RC or contradicted common expectations.

  • Losses and backbones: Dice+top-k CE improved ET and TC by under one point each but degraded RC by more than two, and RC-only routing reproduced the failure.The study therefore could not compensate for RC rarity through the loss alone at this dataset scale.
  • Losses and backbones: Focal-Tversky variants failed: γ = 0.75 collapsed ET and RC Dice to zero within 100 epochs, while a milder setting never exceeded the Dice+CE baseline by epoch 613.A MedNeXt-L sixth ensemble member also degraded avg4 under both tested weightings.
  • Post-processing and inference: Removing RC components below 100 or 300 voxels degraded RC by up to seven points, while ungated five-iteration ring dilation caused the largest RC regression by extending into skull and CSF.These failures motivated the conservative, gated i = 3 BoundaryExpand design.
  • Post-processing and inference: Lowering the RC softmax threshold to 0.40 traded precision for recall at a net loss, and an SNFH < 10 mm3 filter was neutral and dropped.Four region-swap fusion variants also produced no gain after correcting an axial I/O bug.
  • Post-processing and inference: A 0.25 sliding-window sweep with 75% overlap was consistently worse than 0.5, yielding avg4 0.6795–0.6796 versus 0.6823.The authors attribute this result to over-averaging at patch boundaries.

5 Discussion

Under lesion-wise Dice, post-processing is asymmetric: recovering missed lesion extent is safer than deleting components, making metric-aware operations consequential.

  • Recovering partial lesion volume is safer under LW-DSC than deleting a component, because deletion scores that lesion zero.For a 200-voxel lesion, expanding an 80-voxel prediction to 160 raises Dice from 0.57 to 0.89, while deleting the component scores 0.
  • Confidence-gated BoundaryExpand recovers resection-cavity recall with sign-consistent gains without harming the other channels.
Loading 2609.11477v1…