Auditing Synthetic Tumor Boundaries with Shift-Robust Haar Regularity

Synthetic lesions pasted into healthy CT scans are now good enough to replace scarce real validation data. Synthetic Data as Validation raised out-domain liver-tumor Dice on FLARE'23 from 31.1% to 35.4% this way. Nobody, however, checks whether those lesions are rough in the right places. The generator builds tumors from deformed ellipsoids, Gaussian noise and Gaussian smoothing. That recipe can produce interiors that are too smooth and rims that are either too blurry or show a hard voxel-grid step. A segmentation network can learn to find tumors by detecting these artifacts rather than the biology. A validation set built from the same generator will then reward that shortcut instead of exposing it.

We propose measuring local roughness directly. The quantity is the pointwise Hölder exponent $h(x)$. It says how sharply intensity changes near a point: $h$ near 0 means a jump, and $h$ near 1 means nearly smooth. We compare its distribution in a thin band around synthetic tumor boundaries against the same distribution around real annotated tumors. The estimator is the redundant Haar frame of Regularity properties of Haar Frames. A single set of step-function "Haar" wavelets misreads points that sit on its grid boundaries. The union of three copies shifted by 0, 1/3 and 2/3 removes that blind spot, and it recovers $h(x)$ everywhere through "wavelet leaders", which are local maxima of coefficients across scales. The claim is that this regularity gap is an interpretable, localized realism score for synthetic lesions. We will test whether it predicts which generator settings actually help on real external scans, and whether it can steer the generator.

Core Claim

A synthetic-lesion generator should be judged by how well its interface-band exponent distribution matches that of real lesions, and not only by downstream accuracy after the fact. Write $P_{\text{syn}}(h)$ and $P_{\text{real}}(h)$ for the empirical distributions of $h$ estimated within a few millimetres of the synthetic and real tumor boundaries. We define the gap as

$$ G = W_1\big(P_{\text{syn}}(h),\, P_{\text{real}}(h)\big), $$

where $W_1$ is the one-dimensional Wasserstein distance. We compute $G$ separately for the interior, the rim and the surrounding parenchyma. We predict three things. $G$ will be large for the default Gaussian-smoothing generator. Across generator variants, $G$ will rank downstream out-domain performance better than generic realism scores do. Variants with large rim-band gaps will let segmenters use a shortcut we can detect.

The theory sets the scope. Jaffard and Krim's characterization holds only for $0<\alpha<1$, because a Haar element cancels only constants and has a single vanishing moment. That range covers what matters at a lesion rim, which runs from jump-like to mildly rough. It cannot distinguish smooth from very smooth texture. Near-white CT noise also sits below this range. For that reason we estimate on lightly denoised volumes and report the scale range used. We do not claim the method sees smoothness above 1. The theory is also stated on the real line. We therefore sample one-dimensional intensity profiles along boundary normals and tangents, so the proven one-dimensional result applies directly, and we avoid an unproven three-dimensional extension.

Relation to Prior Work

Two papers motivate the gap. Hu et al. show that synthetic validation stabilizes checkpoint selection and improves tiny-tumor sensitivity from 33.9% to 52.3% out-domain. Their texture and blending are heuristic, and they never measure boundary regularity; we audit and adjust their generator rather than replace it. On the Equivalency, Substitutability, and Flexibility of Synthetic Data measures a roughly 30:1 synthetic-to-real exchange rate for tracking. It treats the generator as a black box and explains success only through behavioral labels. We aim to open the box, in a different modality, with a measurable signal-level property.

Three papers supply the method or constrain it. Jaffard and Krim provide the estimator but no experiments, so part of our contribution is the first numerical validation of the three-shift Haar frame on signals with known exponents. The Regularity of Refinable Functions proves that compactly supported refinable functions built from finite filters cannot be infinitely smooth. Moving to smoother wavelets therefore trades away locality for more vanishing moments, and this is why we test the trade explicitly rather than assume it. Fast Random Access to Wavelet Compressed Volumetric Data Using Hashing documents grid-aligned blocky artifacts in Haar-coded CT volumes. This is the same kind of lattice-aligned step we suspect synthetic insertion creates, and it is the case a single unshifted Haar basis handles worst.

Research Program

The first stage calibrates the estimator. We build one-dimensional signals with known exponents, including fractional Brownian paths with prescribed roughness, single jumps placed at dyadic and non-dyadic positions, and cusps. On these we measure the bias and variance of $h$ estimated by three methods: a single Haar basis, the three-shift Haar frame, and Daubechies-4 wavelet leaders. The same checks are repeated after adding CT-like noise and smoothing.

The second stage is the audit. We extract normal and tangential profiles around real tumors in LiTS, using its annotations, and around synthetic tumors generated with the Hu et al. pipeline on the 75 healthy CHAOS, BTCV and Pancreas-CT volumes. From these we compute band-wise $G$.

The third stage builds a family of roughly twelve generator variants. They vary smoothing width, texture type (Gaussian versus fractional noise with a set roughness), rim blending (hard composite, Gaussian feather, or sub-voxel anti-aliased insertion) and capsule modeling. For each variant we train and validate segmenters exactly as Hu et al. do and record FLARE'23 Dice and tiny-tumor sensitivity.

The fourth stage closes the loop. We tune the generator's roughness parameters to minimize $G$ and test whether the result beats the default.

Baselines

The naive reference is real-only validation, meaning model selection on a small static real split. Hu et al. report out-domain FLARE'23 Dice of 31.1% and tiny-tumor sensitivity of 33.9% for this setting. The strongest existing method is Hu et al.'s default dynamic synthetic validation, reported at 35.4% FLARE'23 Dice and 52.3% tiny-tumor sensitivity. We rerun it with our own seeds so that data, architecture, training schedule and compute are identical across all variants. The regularity-matched generator counts as beating it if it improves FLARE'23 Dice or tiny-tumor sensitivity beyond the seed-to-seed spread across at least three seeds, with no change in real data or compute.

For the claim that $G$ predicts usefulness, the competing predictors are two generic scores. The first is a learned realism score: a 3D classifier trained to tell real from synthetic lesion patches, with its accuracy used as the gap. The second is a hand-crafted edge statistic: mean intensity-gradient magnitude and its variance across the boundary. Each predictor is scored by Spearman correlation with FLARE'23 Dice across the variant family, using the same patches. We have no reported result on Lacuna for either predictor on this task.

Two ablations isolate the key ingredient, the shift redundancy. The first replaces the three-shift frame with a single Haar basis. The second replaces it with Daubechies-4 leaders at matched scale ranges.

Why not just…?

Why not just train a real-versus-synthetic classifier and use its accuracy as the realism score? This is the most serious threat. A classifier may predict downstream usefulness as well as $G$ does or better. Its weakness is that it gives a single number that does not say what to change. It can also latch onto global intensity cues unrelated to the boundary. The correlation comparison across generator variants settles which predictor ranks better. If the classifier wins, the remaining contribution of our method is diagnostic: it tells the generator designer which band, and which direction of roughness, is wrong. We would report that narrower claim honestly.

Why not just use smooth wavelets such as Daubechies, which have more vanishing moments? They extend the measurable range above 1 and are the standard tool for wavelet-leader analysis. Wider filters, however, blur the few-voxel band where insertion artifacts live. By the refinable-function bound, more smoothness always costs more support. Lattice-aligned steps are exactly where shift sensitivity matters most. Even so, Daubechies leaders may perform just as well. The calibration stage on jumps placed at dyadic positions, and the ablation on predictive correlation, decide the question directly. If Daubechies leaders tie, we adopt them and credit the regularity-profile idea rather than the Haar construction.

Why not just skip auditing and pick generator settings by real out-domain validation? That uses exactly the scarce real annotations that synthetic validation exists to spare, and Hu et al. show small real splits overfit. Our loop needs only real tumor boundaries for the reference distribution, not a held-out outcome set. We will also report the oracle that selects variants using FLARE'23 itself. This shows how close the $G$-based selection comes to that upper bound.

Decisive Checks

On the calibration signals, the three-shift frame must recover known exponents at dyadic jump positions where the single Haar basis is biased. If it fails, the method is unsound.

The default Hu et al. generator must show a rim-band gap clearly larger than the spread between two disjoint halves of the real LiTS tumors. If synthetic and real are already indistinguishable at that noise floor, the audit has nothing to fix.

For the shortcut test, we take segmenters trained on the most mismatched variant. We evaluate them on held-out synthetic tumors whose rims are re-rendered to match real roughness while the shape stays fixed. A marked drop in Dice relative to unmodified synthetic tumors would show the network relied on the artifact.

Finally, $G$-minimized generation must match or exceed the reproduced default on FLARE'23 Dice and tiny-tumor sensitivity. The Spearman correlation of $G$ with out-domain Dice must also be at least that of the classifier score. If neither condition holds, the regularity gap is at best a descriptive tool.