When Fuzzy Tumor Borders Make Risk Interactions Lie

Clinical risk models increasingly use shape features measured from 3D scans, such as tumor volume, surface area and sphericity. Each of these depends on where a segmentation draws the tumor's edge. In low-contrast CT that edge is uncertain: two radiologists, or two plausible model outputs, can disagree by several voxels all around the lesion. Non-linear risk models such as random survival forests, boosted trees and small neural networks are prized because they capture interactions between factors. An interaction, though, is a product of two measurements, and multiplying two noisy measurements compounds their noise. With only a few hundred patients, a flexible model fit to boundary-dependent features may learn interactions that are mostly segmentation noise. Under the same conditions, a plain additive Cox model, which estimates hazard as a weighted sum of factors, would look more robust.

This note proposes to measure that effect and then correct for it. A 3D segmenter is trained on scans annotated by several radiologists. It is used to produce, for every patient, a set of plausible tumor masks, and from these a per-patient distribution of each shape feature. The risk model is then fit so that each interaction is trusted only in proportion to how reliably its component features can be measured. The core prediction is quantitative: the value of non-linear interactions in small, low-contrast cohorts is set by the product of the boundary-induced reliabilities of the interacting features. If this holds, it explains why simple models often win in this regime, and it offers a way to keep the interactions that are real.

Core Claim

Classical measurement-error theory gives the starting point. Write a shape feature as $x_j$. Its between-patient variance is $\sigma_j^2$, and the within-patient variance across plausible segmentations is $\tau_j^2$. Its reliability is

$$\lambda_j = \frac{\sigma_j^2}{\sigma_j^2 + \tau_j^2}.$$

For centered features with independent errors, the observed product $x_j x_k$ has reliability $\lambda_j \lambda_k$. The ratio of true to observed variance factorizes because the expectations of products of independent zero-mean terms factorize. A pair of features that are each 70% reliable therefore yields an interaction only about 49% reliable. Main effects lose signal linearly in $\lambda$, while interactions lose it multiplicatively. Low-contrast boundaries push $\tau_j$ up, and small samples prevent a flexible model from averaging that noise out.

The claim has two parts. The first is that the advantage of non-linear over additive risk models on imaging-derived features should track the estimated $\lambda_j\lambda_k$ of the interactions those models exploit. The second is that a risk model which (a) integrates over sampled masks during training and prediction and (b) shrinks each interaction by its estimated reliability will recover the interaction benefit that deterministic-mask pipelines lose.

Prior work splits along exactly this seam. Automated Segmentation and Recurrence Risk Prediction of Surgically Resected Lung Tumors with Adaptive Convolutional Neural Networks segments lung tumors with a 2D residual U-Net (70.3% Dice on LIDC-IDRI). It then crops a fixed 3D cube and feeds deep features plus clinical covariates to a random forest, passing no information about edge uncertainty forward. That work motivates the proposal, because errors from a single mask flow straight into the interaction model. Survival prediction of NLST lung cancer patients from radiomics and clinical data uses radiologist-supervised masks on 263 adenocarcinoma patients. It finds that an additive Cox model with clinical and shape features beat random survival forests. That result constrains the proposal: it is the strongest evidence that non-linearity can hurt here, and this proposal offers an explanation that can be tested rather than an argument against it. Beyond Boundary Noise shows that voxel-wise entropy concentrates on fuzzy edges and misses cases where experts disagree whether a nodule exists at all. Its post-hoc ambiguity head recovers that missing signal, but it never connects uncertainty to patient outcomes. It supplies both a method we reuse and a warning about which uncertainty summaries to trust.

Research Program

The first component is a source of plausible masks. A native 3D segmenter is trained on LIDC-IDRI, which provides up to four independent radiologist contours per nodule. The work uses 3D rather than slice-wise processing because inter-slice context is what resolves low-contrast edges (3D Deeply Supervised Network for Automatic Liver Segmentation from CT Volumes), and because 3D shape features are the target. To make the output a distribution rather than a single mask, we train a small ensemble in which each member is supervised by a different annotator's contours. This is the simplest construction that targets the spread between experts directly, instead of the spread between dropout samples. The pre-train-then-adapt protocol follows A Neighborhood Attention Transformer Network for Enhanced 3D Segmentation of the Left Anterior Descending Artery. That work showed that, for a faint structure on 20 target scans, large-source pre-training plus low-rank adaptation was needed. Here the segmenter is adapted to NLST low-dose CT with low-rank adapters (small trainable weight updates added to frozen layers).

The second component turns masks into feature distributions. For each patient we draw $K = 20$ masks and compute 3D shape features from each. The features are volume, surface area, sphericity, elongation and maximum diameter. As a reference we also compute the 2D maximum diameter emphasized by Leu et al. From the spread of these features across masks we estimate each patient's $\tau_j^2$, and from the cohort we estimate $\lambda_j$. We also run the Beyond Boundary Noise ambiguity head on the frozen segmenter bottleneck. Its score is used only as a diagnostic, to flag patients whose lesion extent is contested in a way the mask spread may understate.

The third component is a reliability-aware risk model. We use a Cox model with an additive part plus pairwise interactions among clinical variables (age, stage, smoking history) and shape features. It is fit by averaging the partial likelihood over each patient's $K$ masks. Each interaction coefficient $\beta_{jk}$ receives a ridge penalty proportional to $1/(\lambda_j\lambda_k)$; clinical variables are treated as reliability 1. A boosted survival model trained on the mask-augmented data, with predictions averaged over masks at test time, serves as the tree-based counterpart. This keeps the comparison from depending on one model family.

The fourth component is an analysis of when interactions help. Cohorts are subsampled to 50, 100 and the full available size. For each size we estimate how far the non-linear model's C-index exceeds the additive model's, and we test whether that gap correlates with the mean $\lambda_j\lambda_k$ of the interactions the model actually uses.

Baselines

The naive reference point is a Cox model on clinical variables alone. On the NLST adenocarcinoma cohort, Leu et al. report a C-index of 54.81 for clinical data alone. On NLST recurrence, Basta et al. report a clinical-only AUC of 56% with 0% recall of high-risk Stage I patients. Any imaging-based method must clear these by a wide margin to be worth its cost.

The strongest additive competitor is Leu et al.'s Cox model on clinical plus shape features from a single mask. They report a C-index of 67.06 and a mean time-dependent AUC of 71.27. We re-run it on our reconstructed cohort, with features from the consensus mask of our segmenter and with the same cross-validation folds and the same hyperparameter search budget as every other model. We do this because we cannot guarantee an identical patient list. The cited figure is a sanity reference, not the comparison target.

The strongest non-linear competitor is the Basta et al. pipeline: a deterministic mask, a fixed cube crop, 3D CNN features, and clinical covariates fed to a random forest. They report a recurrence AUC of 73.0% and an accuracy of 78%. We reimplement it with our 3D segmenter in place of their 2D one, so that the comparison isolates uncertainty handling rather than segmenter quality.

Two ablations isolate the key ingredient. In the first, masks are sampled and predictions averaged at test time only, with no likelihood integration and no reliability shrinkage. In the second, likelihood integration is kept but every interaction gets a uniform penalty, so that we can see whether the $1/(\lambda_j\lambda_k)$ scaling matters.

Beating the field means three things. On held-out patients, the full method must exceed the best of these baselines in C-index. It must do so by more than the bootstrap standard error at the 100-patient subsample size. It must also not lose to the additive Cox model at any sample size.

Why not just…?

Why not just use the additive Cox model with clinical and shape features, as Leu et al. found best? This is a real threat, and the proposal is built so that it can lose to this model. If the reliability-aware Cox model gains only through its additive part, with interaction coefficients shrunk to near zero, we report that interactions among boundary-derived features are not recoverable at these sample sizes. That outcome would still be a useful, mechanistic result. The proposal survives as a method only if the interactions it retains add a measurable gain on held-out data. Mask integration alone may also improve the additive Cox model, and we will report that separately.

Why not just append mean voxel entropy as an extra risk feature? Beyond Boundary Noise found that aggregated entropy reached an AUROC of only 0.48 on LIDC-IDRI and 0.54 on LNDb for detecting expert disagreement, against 0.74 and 0.67 for the ambiguity head. A single scalar also cannot tell the model which feature is unreliable. A blurry edge distorts surface area far more than volume. We include "Cox plus mean entropy" as a baseline, so this argument is tested rather than assumed.

Why not skip segmentation and learn directly from an image cube or voxel-level feature maps? Mask-free models avoid boundary error by construction. Cross-dataset adaptation of voxel-level deep radiomics reaches a C-index of 0.73 on 150 external patients after fine-tuning on 37, though in a different disease setting and with pseudo-3D slicing. This is a genuine competitor. We therefore include a mask-free 3D baseline, a cube-crop 3D CNN with clinical covariates, trained under the same transfer protocol. If it matches our method, then explicit shape features are not worth their boundary cost on NLST. Leu et al.'s finding that shape outperformed texture in small cohorts makes us expect the opposite, but the experiment decides.

Why not just average predictions over several masks at test time? That is our first ablation. Theory suggests it is insufficient: test-time averaging reduces the variance of the prediction, but it does nothing about interaction coefficients that were inflated by noisy features during training.

Proposed Experiments

The first decisive check is a controlled test of the attenuation law on LIDC-IDRI, where ground-truth measurement error is observable. We compute shape features from the multi-annotator consensus and treat them as "true." We then simulate survival outcomes from a known model with specified main effects and one shape-by-clinical interaction. Next, we fit additive and interaction models to features computed from single-annotator masks and from our sampled masks. The prediction is that the recovered interaction coefficient shrinks by roughly $\lambda_j\lambda_k$, while main effects shrink by roughly $\lambda_j$, and that reliability-aware fitting reduces the bias. If the observed attenuation does not follow the product form, the core claim fails.

The second experiment is the main evaluation on NLST. It covers survival in the adenocarcinoma cohort reconstructed from Leu et al.'s criteria and recurrence following Basta et al.'s setup. All baselines above are run at subsample sizes of 50, 100 and the full cohort. The metrics are C-index, time-dependent AUC and high-risk Stage I recall. We also report how the non-linear-versus-additive gap relates to mean interaction reliability across sizes and resampling seeds. A secondary analysis compares the reliability of 3D shape features with Leu et al.'s 2D maximum diameter. It asks whether volumetric shape is less boundary-sensitive, and whether that carries over to risk prediction. It also asks whether removing the patients flagged by the ambiguity head changes the ranking of methods.

The third experiment is a replication in a second organ. TCGA-KIRC provides 156 renal cancer patients with preoperative 3D CT and recurrence follow-up. There, Boeke et al. (Integrating Pathology and CT Imaging for Personalized Recurrence Risk Prediction in Renal Cancer) report a Leibovich clinical score C-index of 0.805 raw and 0.749 with random tie-breaking, against 0.775 for their deep fusion model. The dataset has no multi-annotator contours, so uncertainty comes from an ensemble adapted from the lung segmenter. The test is the weaker one: whether reliability-aware interaction fitting improves on single-mask fitting, measured against the tie-broken Leibovich score. We state plainly that this transfer from lung nodules to renal masses may fail, and we will report it either way.