A radiomics nomogram is a point-based chart that doctors use to estimate a patient's risk. It adds two things together: a handful of routine clinical variables, such as age, performance status and tumor size, and a "Rad-score," a single number compressed from hundreds or thousands of texture and shape measurements extracted from a scan. Published nomograms report strong discrimination, but the radiomic half has well-known failure modes. Its features shift with scanner and reconstruction settings, it is selected on cohorts of one to two hundred patients, and many of its features silently re-encode clinical facts such as tumor volume. When such a nomogram reaches a new hospital, the radiomic term can drag a sound clinical estimate in the wrong direction, and nothing in the standard pipeline limits how far.
Cellular networks faced the same structure and solved it explicitly. A dense layer of small cells shares spectrum with a trusted, high-priority macro layer. The rule that makes this work is that the secondary layer may transmit only in directions that do not hit the primary user, and only up to a total interference budget the primary layer sets and enforces. We propose to carry both halves of that rule into nomogram construction. First, the Rad-score is built only from the part of the radiomic features that the clinical variables cannot already explain. Second, its contribution to each patient's log-odds is capped by a budget chosen on held-out data. The goal is a nomogram that is never much worse than the clinical model, even under scanner shift, and that keeps most of the radiomic gain when the imaging is in-distribution.
Core Claim
The claim is empirical and falsifiable. Across external-center or shifted-acquisition validation, a nomogram whose radiomic term is (a) orthogonalized against the clinical variables and (b) budget-limited will have a smaller worst-case loss relative to the clinical-only model, and better calibration, than the standard concatenate-then-LASSO nomogram. It will also give up little discrimination in-distribution. A secondary claim is that the orthogonalization step makes radiomic feature selection more stable across bootstrap resamples, because shape features that stand in for tumor size stop competing with the clinical size variable.
Let $c$ be a patient's clinical vector, $r$ their standardized radiomic vector, and $P_C$ the least-squares projection of radiomic features onto the span of the clinical variables, estimated within each training fold. The model is
$$\text{logit}\, p = f(c) + \text{clip}_{[-Q,\,Q]}\big(g(r_\perp)\big), \qquad r_\perp = (I - P_C)\, r,$$
where $f$ is the clinical logistic or Cox model, $g$ is a sparse linear Rad-score, and $Q$ is the maximum log-odds (or log-hazard) shift the radiomic term may impose on any patient. During training the hard clip is replaced by a penalty $\lambda \cdot \mathbb{E}[\max(0, |g| - Q)^2]$. In this penalty, $\lambda$ plays the role of the interference price in the leader–follower game: the clinical model "charges" the radiomic model for every unit of shift beyond the budget.
Where the Borrowed Ideas Come From
The design follows three results from the network literature. In cognitive interference alignment, Maso et al. precode small-cell transmissions into the null space of the macro user's channel, so cross-tier interference is zero by construction. Our residualization $r_\perp$ is the statistical analogue: the radiomic tier may carry only information orthogonal to the clinical span. The Stackelberg formulation of Kang et al. has the macro station cap aggregate femtocell interference at a threshold $Q$ through a price, solved by bisection on aggregate measurements alone. We reuse the threshold-plus-price structure and the bisection search over $\lambda$. The eICIC study of López-Pérez et al. is the cautionary motivation. Blanking subframes took macro-user outages from 267 to 0 at roughly a 25% cost in femtocell throughput, which shows that protecting the primary tier has a measurable price on the secondary one. We expect the same trade, a small in-distribution AUC cost for a large cut in worst-case harm, and we will report it as such.
Work that literally combines tiers and healthcare, such as cross-silo federated learning over multi-tier networks, FedDCT and FLamby, treats "tier" as a network or institutional hierarchy and studies convergence and communication. We borrow none of that machinery. Our "tiers" are information sources inside a single predictor.
On the radiomics side, the nomograms of Huang et al. for spread through air spaces in lung adenocarcinoma, the TERT-mutation glioblastoma nomogram, and Wenwen et al. for triple-negative breast cancer survival all fuse the two tiers by selecting radiomic features and summing linear terms. None of them bounds the radiomic term or removes its clinical overlap. All three name scanner dependence as their main limitation, and that limitation is what our method targets. Parmar et al. constrain our design in two ways. They find that unsupervised feature reduction (PCA) beat supervised filtering at N=112, with mean AUC 0.70 vs 0.67, so our follower model $g$ will include a PCA-on-residuals variant. They also attribute 65% of performance variance to the endpoint itself, a warning that method effects will be small and need paired statistics.
Research Program
We will build an open implementation that wraps existing pipelines. Features come from PyRadiomics or the CaPTk features shipped with the data. The projection $P_C$ is estimated inside each training fold, never on the test data. The budget $Q$ is chosen by bisection on an inner validation split against a target: the largest $Q$ whose worst inner-fold loss relative to the clinical model stays below a tolerance fixed in advance.
The analysis has three parts. The first measures "cross-tier leakage": the fraction of each selected radiomic feature's variance explained by clinical variables, reported for both the published-style and the orthogonalized nomograms. The second measures selection stability, as the Jaccard overlap of selected features across 200 bootstrap resamples. The third, the central test, measures harm under shift.
Data and Shift Settings
The primary public resource is UPenn-GBM, which provides 630 glioblastoma patients with clinical variables, survival, IDH status (515 cases) and MGMT status (317 cases), plus CaPTk radiomics for three tumor subregions. We will build survival and IDH and MGMT nomograms with age as the core clinical tier. The file reports no predictive result on this cohort, so all of its baselines will be ours. Shift will be created two ways. The first splits by acquisition parameters wherever the imaging headers allow. The second re-extracts features after controlled resampling, intensity normalization and kernel-like smoothing changes that mimic reconstruction differences.
A cross-cohort test trains on the BraTS 2019 high-grade glioma survival subset used by Yousaf et al. and tests on UPenn-GBM, after removing patients that appear in both. We will also seek collaborative access to the dual-center spread-through-air-spaces cohort and the breast-cancer cohort with its prospective external validation set. These are the settings where a genuine second center exists. We state plainly that this access is uncertain, and the glioma experiments are designed to be decisive without it.
Baselines
The naive reference is the clinical-only model: logistic or Cox regression on age, performance status where available, and tumor size. It defines the "primary tier" our method promises not to fall far below, and every result is reported as a paired difference against it.
The strongest existing method is the standard published nomogram: concatenate standardized clinical and radiomic features, select with LASSO, form the Rad-score, and refit jointly. This is the pipeline of Huang et al., which reports AUC 0.848 on its external center, and of the TERT nomogram, which reports validation AUC 0.899. Wenwen et al. report external AUC between 0.73 and 0.89. These numbers come from cohorts we may not obtain, so on UPenn-GBM we re-implement the pipeline exactly and give it the same folds, the same features and the same nested tuning budget for its penalty that our method receives for $\lambda$ and $Q$. For a nonlinear ceiling we add a random forest on PCA-reduced features, following Parmar et al., who report best AUCs of 0.76 for recurrence and 0.77 for death. On BraTS-style three-class survival, Yousaf et al. report 75% patient-level accuracy for a random forest.
Beating the standard nomogram means three things together under shift. Its worst-case drop in AUC relative to the clinical model must be smaller, with the difference's bootstrap confidence interval excluding zero. Calibration error must be lower. In-distribution AUC must be within 0.02 of the standard nomogram. Two ablations isolate the ingredients: one keeps orthogonalization without the budget, and the other keeps the budget without orthogonalization.
Why not just…?
Why not just leave the clinical variables unpenalized inside the LASSO? This is the most serious threat to the orthogonalization half. For squared-error loss, profiling out unpenalized clinical coefficients is mathematically equivalent to residualizing the radiomic features. For logistic and Cox loss the equivalence is only approximate. We therefore expect the "orthogonalization only" ablation to roughly match this baseline. If it does, we will say so and credit any gains to the budget alone. Orthogonalization would then remain useful mainly as a diagnostic, because the leakage measurements make the Rad-score's incremental meaning explicit.
Why not just recalibrate the Rad-score coefficient at the new site? Recalibration needs labeled outcomes from the new center. A nomogram arriving at a new clinic usually has none, and our method is meant for that zero-label deployment. We will still run site-recalibration with 10, 25 and 50 local labeled patients as an upper reference, to show how much of its benefit the budget recovers without labels.
Why not just winsorize the Rad-score after fitting? This is a real threat to the budget half, because the hard clip at deployment is a winsorization. The difference is that our follower model is trained knowing the budget exists, so it allocates its limited shift to the patients where radiomics is most informative. We will test post hoc clipping at the same $Q$ directly. If it matches training-time budgeting, the contribution shrinks to a principled way of choosing $Q$ plus the empirical case for capping, and we will report it that way.
Why not just use a random forest or an end-to-end network? Neither produces the point chart clinicians read, and neither bounds harm under shift. The small-sample evidence is mixed: Yousaf et al. saw a fully connected network fall to 51.5% accuracy against 76.5% for a random forest on hand-crafted features, while the PET/CT deep radiomics study reached AUC 0.81–0.86 with larger multi-institution data. We include the forest as a ceiling on in-distribution accuracy, not as the target.
Decisive Checks
The proposal succeeds only if the budgeted nomogram shows a smaller worst-case shortfall against the clinical model than the standard nomogram, across the shift settings on UPenn-GBM and in the BraTS-to-UPenn transfer. The comparison uses paired bootstrap at matched in-distribution AUC.
It fails if either of two things happens. One failure is that standard LASSO nomograms show no meaningful harm under realistic shift. In that case the problem we target is not real at this scale, and that finding is worth publishing. The other failure is that post hoc clipping or unpenalized clinical covariates reproduce every gain, in which case the network-derived machinery adds nothing.
A third check tests whether the leakage analysis matters. It should show that a substantial share of selected features in published-style nomograms are largely explained by tumor size or age, and that removing this overlap raises bootstrap selection stability. If neither appears, the "cross-tier interference" story for radiomics is wrong.
References
- A Distributed Approach to Interference Alignment in OFDM-based Two-tiered Networks
- Price-Based Resource Allocation for Spectrum-Sharing Femtocell Networks: A Stackelberg Game Approach
- Enhanced intercell interference coordination challenges in heterogeneous networks
- Cross-Silo Federated Learning for Multi-Tier Networks with Vertical and Horizontal Data Partitioning
- FedDCT: A Dynamic Cross-Tier Federated Learning Framework in Wireless Networks
- FLamby: Datasets and Benchmarks for Cross-Silo Federated Learning in Realistic Healthcare Settings
- Dual-layer spectral detector CT quantitative parameters and radiomics for predicting spread through air spaces of lung adenocarcinoma: a dual-center study
- A radiomics‐based nomogram may be useful for predicting telomerase reverse transcriptase promoter mutation status in adult glioblastoma
- Integrating ultrasound radiomics and clinicopathological features for machine learning-based survival prediction in patients with nonmetastatic triple-negative breast cancer
- Radiomics-based Prognosis Analysis for Non-Small Cell Lung Cancer
- The University of Pennsylvania glioblastoma (UPenn-GBM) cohort: advanced MRI, clinical, genomics, & radiomics
- Brain Tumor Survival Prediction Using Radiomics Features
- Non-invasive decision support for NSCLC treatment using PET/CT radiomics