Distance Map Auxiliary Loss for Brain Tumor Segmentation

We study the use of a Signed Distance Transform (SDT) auxiliary loss on top of MedNeXt-B / nnU-Net v2 for 3D brain tumor segmentation on BraTS 2023 GLI. The SDT task brings a modest but statistically robust Dice improvement (+0.63 to +0.96 pp per region, Wilcoxon p < 10⁻⁶, n = 240) while introducing a new failure mode: spurious isolated connected components (fragments) not present in the ground truth, most acute in NCR and ED.

We propose a parameter-free post-hoc connected-component consensus filter (CC-consensus filter) : start from the DistMap prediction and, for each class, drop any connected component whose same-class mask has zero voxel overlap with the Baseline prediction. This rule is not a Mixture-of-Experts in the strict gating-network sense ; it is a hard-label consensus filter at the connected-component level. It reduces NCR fragments by 81 % (Wilcoxon p = 7 × 10⁻⁴) at no Dice cost.

A large-scale ceiling analysis on 1196 patients shows that the oracle per-class selection upper-bound is only +0.005 Dice avg above the default CC-consensus rule; 4 classifier families trained on 31 hand-crafted features (tumor morphology, topology, inter-model agreement) fail to robustly beat the default CC-consensus in 5-fold cross-validation. The gap cannot be closed without voxel-level probabilistic voting or architectural diversity — motivating a training-time fragment-aware loss over further post-hoc engineering.

import numpy as np
from scipy import ndimage as ndi

STRUCT_26 = ndi.generate_binary_structure(3, 3)

def cc_consensus_filter(distmap_seg, baseline_seg, classes=(1, 2, 3)):
    """Connected-component consensus filter.
    Start from DistMap; for each class, drop any CC whose same-class mask has
    zero voxel overlap with Baseline. Parameter-free (26-connectivity)."""
    filtered = distmap_seg.copy()
    for c in classes:
        d_mask = (distmap_seg == c); b_mask = (baseline_seg == c)
        if not d_mask.any(): continue
        lab, n = ndi.label(d_mask, structure=STRUCT_26)
        for cc_id in range(1, n + 1):
            cc = (lab == cc_id)
            if not np.any(cc & b_mask):
                filtered[cc] = 0  # unconfirmed fragment
    return filtered

All numbers in this paper are reproducible from the GitHub repo in under 10 min on a modern CPU.

• scripts/extract_patient_features.py — 20 GT-morphology features per patient (~5 min, 10 E-cores)
• scripts/extract_agreement_features.py — 11 inter-model agreement features (~6 min)
• scripts/select_demo_patients.py — 6 champions (C1–C6)
• scripts/oracle_per_class.py — patient and per-class oracle bounds
• scripts/sweep_adaptive_fusion.py — threshold sweep (~2 min, 10 workers)
• scripts/meta_selector_perregion.py — 3 classifier families × 5-fold CV
• scripts/simple_rule_cv.py — 1-feature rule in 5-fold CV (overfit check)

@article{cassez2026ccconsensus,
  title   = {Distance Map Auxiliary Loss for Brain Tumor Segmentation:
             A Fragment-Centric Analysis and the Saturation Ceiling of
             Post-hoc Connected-Component Consensus Filtering},
  author  = {Cassez, Guillaume},
  journal = {arXiv preprint},
  year    = {2026},
  url     = {https://guillaume-cassez.fr/brats/paper1/},
  doi     = {10.5281/zenodo.19695264}
}

Feedback welcome at cassez.guillaume@gmail.com, via the GitHub issues, or on Bluesky @guillaume-cassez.bsky.social.