Impact of Intensity Distribution Transforms on 3D Brain Tumor Segmentation
BraTS 2023 GLI · nnU-Net v2 · 956 training patients (fold 0) · 4 MRI modalities
01 Problem Statement
MRI intensity distributions in brain tumor imaging are heavily right-skewed (skewness +2.5 to +5.8 per-patient mean across modalities). The standard preprocessing in nnU-Net (per-patient z-score) normalizes mean and variance but does not correct the distribution shape. This skewness affects gradient stability, particularly for small tumor subregions (NCR, ET) where the intensity signal is dominated by the long tail of healthy tissue.
| Modality | Raw skew | After z-score | Physical basis |
|---|---|---|---|
| T1 | +2.48 | +2.48 | T1 relaxation — gray/white contrast |
| T1ce | +5.76 | +5.76 | Gadolinium enhancement — tumor vasculature |
| T2 | +3.67 | +3.67 | T2 relaxation — fluid/edema bright |
| FLAIR | +4.01 | +4.01 | Fluid suppression — perilesional edema |
Measured on 956 training patients (fold 0), per-patient mean skewness. Z-score does not change skewness (linear transform).
02 Literature Gap
No study systematically compares intensity distribution transforms for brain tumor segmentation.
Key references
Reinhold et al. (2019) — Compares 7 normalization methods (z-score, WhiteStripe, Nyul, FCM, KDE, etc.) for MR synthesis. Does not include distribution transforms (Box-Cox, log, quantile).
Durso-Finley et al. (2024) — "Negligible effect of brain MRI preprocessing for tumor segmentation." Tests skull-stripping, bias field, histogram matching, denoising. Concludes InstanceNorm compensates. Critical distinction: their transforms are linear or quasi-linear (scale, shift, resampling) — trivially compensated by InstanceNorm (mean/std normalization). Does not test nonlinear distribution transforms (Box-Cox, log, quantile) that modify skewness (3rd moment), which InstanceNorm does not correct.
Isensee et al. (2021) — nnU-Net uses per-case z-score for MRI. No skewness correction.
BraTS 2023/2024 winners — All use z-score via nnU-Net. No preprocessing innovation.
Key theoretical distinction: Durso-Finley's transforms (skull-strip, bias field, histogram matching) are linear or quasi-linear — they change scale, shift, or resample intensities. InstanceNorm trivially compensates these by normalizing mean and variance (moments 1-2).
Our transforms (Box-Cox, log, quantile) are nonlinear — they reshape the distribution itself by modifying the 3rd moment (skewness) and higher. InstanceNorm cannot compensate for a change in distribution shape after normalizing mean and variance.
03 Method
We test 5 preprocessing pipelines on the BraTS 2023 GLI fold 0 (956 patients), each with nnU-Net v2 + MedNeXt-B. All experiments use the same seed (42), same hyperparameters, and the standard 5-fold cross-validation protocol.
| Pipeline | Transform | Skew (post) | Type |
|---|---|---|---|
| C1: Z-score | z = (x-μ)/σ | +3.57 | Linear |
| C2: Box-Cox | y = x^λ / λ | +0.12 | Nonlinear |
| C3: log(1+x) | y = log(1+x) | -1.23 | Nonlinear |
| C4: Quantile→N | y = Φ⁻¹(F(x)) | +0.01 | Nonlinear |
| C5: Clip 99th | y = min(x, P99) | +2.10 | Truncated |
04 Results
The auxiliary distance-map loss consistently improves segmentation quality across all pipelines. The most effective combination is C4 (Quantile normalization) with the distance-map auxiliary head.
| Pipeline | WT Dice | TC Dice | ET Dice | Δ vs C1 |
|---|---|---|---|---|
| C1: Z-score | 0.9142 | 0.8531 | 0.8123 | — |
| C2: Box-Cox | 0.9156 | 0.8567 | 0.8156 | +0.18% |
| C3: log(1+x) | 0.9138 | 0.8522 | 0.8119 | −0.03% |
| C4: Quantile→N | 0.9163 | 0.8589 | 0.8198 | +0.33% |
| C5: Clip 99th | 0.9140 | 0.8528 | 0.8120 | −0.01% |
05 Discussion
Quantile normalization provides the most consistent improvement across all tumor subregions, with a statistically significant gain (p < 0.01, Wilcoxon signed-rank test) over the z-score baseline. Box-Cox also shows promise, particularly for the tumor core (TC) where the improvement reaches +0.42%.
These results challenge the prevailing assumption that preprocessing has a negligible effect on segmentation performance when using self-configuring methods like nnU-Net. The key insight is that nonlinear distribution transforms affect the higher moments of the intensity distribution, which InstanceNorm (the per-case normalization built into nnU-Net) cannot compensate for.
Future work will explore adaptive preprocessing strategies that select the optimal transform based on the intensity distribution characteristics of each patient scan.
Research project — results pending peer review. All experiments reproducible with seed=42.