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

BraTS 2026 brain-metastasis segmentation must handle heterogeneous pre- and post-treatment cases and lesion-wise scoring that emphasizes small-lesion detection. This paper evaluates a five-fold nnU-Net ensemble with rule-based post-processing and audits each stage out of fold. The analysis finds two stages robust across folds and shows that LW-DSC rewards recovering undersegmented lesions more safely than deleting components.

  • Problem

    BraTS 2026 brain-metastasis segmentation involves heterogeneous pre- and post-treatment cases, rare resection-cavity labels, and LW-DSC sensitivity to small-lesion detection.

    Source: §1 · p. 2

  • Method

    The pipeline combines a five-fold nnU-Net ResEnc-L ensemble with rule-based post-processing and audits each stage through five-fold out-of-fold analysis.

    Source: §1 · p. 2

  • Results

    0.57 to 0.89: for a 200-voxel ground-truth RC lesion, growing an 80-voxel partial prediction to 160 voxels raises lesion Dice, whereas deleting the component scores 0.

    Source: §5 · p. 10

  • Takeaways & Limitations

    Out-of-fold analysis supports RuleClean and BoundaryExpand as robust across folds, but brainF70’s leaderboard gain does not reproduce out of fold.

    Source: §1 · p. 2, §3.2 · p. 7

    The resection cavity remains the sole critical deficit relative to the paper’s tumour-side results.

    Source: §3.1 · p. 7

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 Introduction

The paper presents a 5-fold nnU-Net ResEnc-L pipeline that focuses on inference-time post-processing for lesion-wise Dice and audits each stage out of fold. It motivates this approach with the 2026 brain-metastasis task’s heterogeneous scans, rare resection-cavity labels, and emphasis on small-lesion detection.

1 Introduction

  • Accurate volumetric assessment supports radiosurgery planning and treatment-response monitoring, while manual RANO-BM measurements under-detect small lesions [14,21,28,1].Source: §1 · p. 1
  • BraTS-MET 2026 [20] [17] combines eight-institution scan heterogeneity, a resection-cavity label in only about one in eight training cases, and lesion-wise Dice dominated by detecting lesions under 27 mm3.The setting includes pre- and post-treatment cases, mixing SRI24-registered and native-space scans.Source: §1 · p. 2
  • The pipeline places ET, TC, and WT in the leading tier on the official validation leaderboard, while out-of-fold analysis supports RuleClean and BoundaryExpand but not brainF70.The leaderboard result is from the validation set, not the hidden test or final ranking; the five-fold audit covers all 1,296 training cases without model-training leakage.Source: §1 · p. 2
  • Thirteen unsuccessful attempts span loss engineering, alternative backbones, and inference settings; finer sliding-window overlap and imbalance-oriented losses can degrade performance.These negative results are presented to help future participants avoid revisiting unsuccessful approaches.Source: §1 · p. 2

2 Methods · 2.1 Dataset

The study uses the complete-modality BraTS-MET 2025 training release for the 2026 challenge, comprising pre- and post-treatment brain-metastasis MRI with five-class annotations. Its official validation set is also complete-modality, while local OOF analyses use held-out predictions for all training cases.

2.1 Dataset

  • The BraTS-MET 2025 release for the 2026 challenge contains pre- and post-treatment MRI cases aggregated from eight institutions, with SRI24-registered and native-space acquisitions.The dataset is cited as [17], and SRI24 as.Source: §2.1 · p. 2
  • Each case has T1c, T1n, T2-FLAIR, and T2w modalities and five annotation classes: background, NETC, SNFH, ET, and RC.NETC is non-enhancing tumour core, SNFH is surrounding non-enhancing FLAIR hyperintensity, ET is enhancing tumour, and RC is resection cavity.Source: §2.1 · pp. 2–3
  • TC is defined as NETC ∪ ET, while WT is defined as NETC ∪ SNFH ∪ ET.Source: §2.1 · p. 3
  • All 1,296 training cases include native T2w, stored in nnU-Net’s four-channel order T1c, T1n, T2-FLAIR, and T2w.The pipeline performs no T2w synthesis or zero-imputation.Source: §2.1 · p. 3
  • The official Synapse validation set contains 179 cases with all four modalities, so the hidden-test inference fallback is not used for reported results.Source: §2.1 · p. 3
  • Local OOF ablations use each of the 1,296 training cases’ held-out-fold predictions, independently of the Synapse validation submission budget.Source: §2.1 · p. 3

2.2 Backbone architecture · 2.3 Training

The method uses a planner-configured 3D full-resolution nnU-Net ResEnc-L and trains five independent folds with the standard nnU-Net trainer, compound loss, and augmentation recipe.

2.2 Backbone architecture

  • ResEnc-L uses ResNet-style residual blocks and instance normalization [25], following the nnU-Net Revisited study.The network is a residual encoder-decoder with deep supervision at every resolution, four input channels, and five output channels for background and four foreground classes.Source: §2.2 · p. 3
  • The official ResEnc-L planner automatically sets network hyperparameters from the dataset fingerprint; only the 3d_fullres configuration is trained.No cascade or low-resolution stage is trained.Source: §2.2 · p. 3

2.3 Training

  • Five folds are trained independently for 1,000 epochs using the standard nnUNetTrainer_ResEncL trainer.Training uses SGD with Nesterov momentum µ = 0.99, weight decay 3 × 10−5, and a polynomial learning-rate schedule with η0 = 0.01 and power 0.9.Source: §2.3 · p. 3
  • The objective is the standard nnU-Net Dice-plus-cross-entropy loss with deep supervision at every decoder scale and no loss re-weighting or focal/top-k modification.The Dice formulation follows V-Net.Source: §2.3 · p. 3
  • Training uses the default nnU-Net augmentation recipe and automatic mixed precision (fp16).Augmentations include mirroring, elastic deformation, rotation, scaling, noise, blur, brightness/contrast changes, low-resolution simulation, and gamma correction.Source: §2.3 · p. 3
  • Each fold is trained on 2×NVIDIA RTX A6000 GPUs with 48 GB VRAM, using planner-selected batch = 2 for 3d_fullres.Source: §2.3 · pp. 3–4

2.4 Post-processing cascade

The post-processing cascade runs three independent scripts sequentially on the five-fold ensemble prediction, keeping the stages separate so each stage’s LW-DSC effect can be quantified.

2.4 Post-processing cascade

  • Three independent post-processing scripts run sequentially on the five-fold ensemble prediction, with separate stages enabling measurement of each stage’s LW-DSC effect.Source: §2.4 · p. 4

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

RuleClean applies four ordered, 26-connected component rules to the argmaxed segmentation, removing selected small or anatomically unattached components. In isolation, it improves the five-fold ensemble’s average validation LW-DSC across ET, TC, WT, and RC.

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

  • 0.6823 average LW-DSC across ET / TC / WT / RC, a +0.0305 gain over the un-post-processed ensemble baseline, on the Synapse validation set.This is the result of applying RuleClean in isolation to the raw five-fold ensemble.Source: Stage 1: rule-based clean-up (RuleClean) · p. 4
  • ET-min10 and RC-min10 remove 26-connected components below 10 mm3, a threshold below the 27 mm3 evaluation cut-off that removes tiny spurious predictions.Removed voxels are relabelled to background.Source: Stage 1: rule-based clean-up (RuleClean) · p. 4
  • SNFH-adj2mm retains whole components only when within 2 mm of TC, while NETC-adj2mm applies the same rule using ET as anchor.These rules preserve retained edema interiors and enforce NETC enclosure by enhancing tumour, with the tolerance absorbing slight ET/NETC boundary imprecision.Source: Stage 1: rule-based clean-up (RuleClean) · p. 4

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

BoundaryExpand targets truncated resection-cavity predictions with a conservative, region-protected expansion in native voxel space. The five-fold out-of-fold audit found a small, consistent RC gain without changing ET, TC, or WT.

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

  • Raw RC lesion-wise DSC was 0.367, as the ensemble truncated cavity contours and LW-DSC penalized subset predictions against larger ground-truth lesions.The boundary shrinkage motivated expansion based on observed under-segmentation, not a target physical distance.Source: Stage 2: resection-cavity boundary… · p. 4
  • The method applies background-restricted 3D ring dilation in native voxel space, gated by PRC > 0.10 and constrained by region protection and a 3× per-component volume cap.The probability gate is a permissive safeguard; replacing 0.10 with 0 produced near-identical outputs because adjacent voxels almost always exceed the threshold.Source: Stage 2: resection-cavity boundary… · pp. 4–5
  • BoundaryExpand yielded a small, sign-consistent RC gain on all five out-of-fold folds and left ET/TC/WT untouched; five ungated dilations were catastrophic.The reported aggressive setting used five dilations with no probability gate.Source: Stage 2: resection-cavity boundary… · p. 6

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

brainF70 removes resection-cavity components mostly outside the brain using a T1c-derived mask, but it failed the five-fold out-of-fold audit and is excluded from the recommended cascade.

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

2.5 Inference and ensembling

Inference uses nnU-Net’s standard sliding-window predictor and an arithmetic-mean five-fold ensemble, with folds loaded sequentially to fit the single 24 GB GPU. A deterministic missing-T2w fallback is defined but was not exercised because all reported cases had native T2w.

2.5 Inference and ensembling

  • Five-fold softmax volumes are fused by arithmetic mean after standard 0.5-step Gaussian-weighted sliding-window inference with three-axis mirroring; weighted averaging and a 0.25 step performed worse or gave no improvement.The five fold models are loaded one at a time, with softmax accumulated on disk before fusion, to run within the Task 1 environment’s single 24 GB A10G GPU.Source: §2.5 · p. 6
  • For cases lacking native T2w, the deterministic fallback reuses T2-FLAIR for the _0003 channel, but no reported case required it.The fallback is specified for the hidden test phase; all reported cases provided native T2w.Source: §2.5 · p. 6

3 Results

The submitted validation pipeline includes brainF70 only to reproduce its leaderboard score, while five-fold out-of-fold results support RuleClean and BoundaryExpand, not brainF70.

3 Results

  • Results report LW-DSC for ET, TC, WT, and RC, with avg4 defined as their arithmetic mean and per-region small-instance F1 added where relevant.Regions absent from ground truth are skipped, as on the leaderboard.Source: §3 · p. 6, §3.2 · p. 7

3.1 Main validation results

  • Table 1 summarizes the final pipeline on 179 Synapse validation cases; the submitted configuration includes brainF70 only to reproduce its leaderboard score.The resection cavity remains the sole critical deficit relative to the paper’s tumour-side results.Source: §3.1 · pp. 6–7, Table 1 · p. 7

3.2 Ablation of the post-processing cascade

  • The five-fold OOF audit scores all 1,296 training cases once with a held-out fold’s own model and the official BraTS evaluation code, without training leakage.Because OOF predictions use single models and a different case population, their absolute values are not directly comparable to ensemble leaderboard scores; stage increments are assessed by sign and per-fold consistency.Source: §3.2 · p. 7
  • RuleClean improves NSD, HD95, and large-instance F1 in every region and roughly halves HD95, but lowers small-instance F1, including ET from 0.360 to 0.277.BoundaryExpand improves RC HD95 from 112.3 to 109.5 mm alongside its RC DSC gain.Source: §3.2 · p. 8, Table 2 · p. 9
  • On five-fold OOF predictions, RuleClean increases avg4 by +0.0640 and RC by +0.1244, while BoundaryExpand adds +0.0021 avg4 and +0.0080 RC consistently across folds.BoundaryExpand leaves ET, TC, and WT unchanged; its RC increment is non-negative on every fold.Source: §3.2 · p. 7, Table 2 · p. 9
  • On OOF predictions, brainF70 changes avg4 by −0.0006 and RC by −0.0020, triggers on only 6 of 1,296 cases, and reverses sign across folds.It removed a true false positive on fold 0 but deleted a genuine resection cavity on fold 2, so it is excluded from the recommended cascade.Source: §3.2 · pp. 7–8, Table 2 · p. 9

4 Failed attempts

The failed-attempt audit finds that several loss and post-processing alternatives did not improve performance, with RC especially vulnerable to aggressive filtering or ungated expansion. The OOF analysis also rejects brainF70 despite a leaderboard gain, while lower RC thresholds, region-swap fusion, and finer sliding-window inference provided no benefit.

4 Failed attempts

  • Dice+top-k CE improved ET and TC by under one point each but degraded RC by more than two; MedNeXt-L and both Focal-Tversky variants also failed.Routing only the RC channel from the top-k model reproduced the failure; γ = 0.75 collapsed ET and RC Dice to zero within 100 epochs, while the milder variant did not exceed the DC+CE baseline by epoch 613.Source: §4 · p. 9
  • RC component filters below 100 or 300 voxels degraded RC by up to seven points, while five-iteration ungated ring dilation caused the study’s largest RC regression.Ungated dilation extended into skull and CSF; these failures bracket the conservative, gated i = 3 BoundaryExpand.Source: §4 · p. 9
  • brainF70 helped one leaderboard case but was net-negative and sign-reversing across five OOF folds; an RC threshold of 0.40 lost net, and SNFH < 10 mm3 was neutral.Four region-swap fusion variants added nothing after fixing an axial I/O bug; a 0.25 sliding-window sweep was worse than 0.5 (avg4 0.6795–0.6796 vs 0.6823), attributed to over-averaging at patch boundaries.Source: §4 · p. 10

5 Discussion

The discussion explains why LW-DSC rewards conservative recall recovery but sharply penalizes component deletion. It notes that RC performance remains a major gap and points to training- and detection-side extensions.

Discussion

  • RC LW-DSC remains 0.549, well below the top validation submissions; the authors identify a dedicated RC pathway as their first-priority extension.They conjecture that foreground oversampling, a connected-component loss on RC, or an RC-specialised head may offer higher-yielding levers than further post-processing.Source: §5 · p. 10
  • For a 200-voxel RC lesion, growing an 80-voxel prediction to 160 voxels raises lesion Dice from 0.57 to 0.89, whereas deleting the component scores 0.By contrast, 40 spurious voxels on an otherwise perfect match reduce Dice only from 1.0 to 0.91, illustrating why recall recovery is safer than deletion under LW-DSC.Source: §5 · p. 10
  • BoundaryExpand recovers RC recall with a sign-consistent gain and no cost to other channels, while over-deletion and ungated over-expansion regress RC.The resulting practical rule is to recover recall conservatively under a probability gate and delete components only when they are confidently false positives.Source: §5 · p. 10

Related papers

Loading 2609.11477v1…