Brain tumor prediction models are usually built as if every patient gets every scan. In practice, the "passive" structural MRI sequences, which show anatomy, are always acquired. The "active" functional sequences are optional. Perfusion imaging tracks an injected contrast bolus through tissue, and diffusion tensor imaging (DTI) measures how water moves through tissue. Both cost scan time, may need extra contrast, and are often missing: in the public UPenn-GBM cohort only 79.6% of patients have perfusion and 88.2% have DTI (UPenn-GBM cohort). This note proposes to learn, from the structural scan and clinical record alone, which individual patients' predictions would change most if functional imaging were added. Functional imaging would then be used only for that subset, and the question is how much of its full-cohort benefit survives when only a fraction of patients receive it.
The design comes from wireless engineering. There, hybrid systems pair a cheap passive reflecting surface with a costly active amplifier, and they spend the active resource only where the passive one fails. The relay in Coexisting Passive RIS and Active Relay-Assisted NOMA Systems serves only the weak, distant user in its hybrid protocol. The selective method in Three High-Rate Beamforming Methods for Active IRS-Aided Wireless Network switches on only the strongest half of the amplifying elements and loses almost nothing. The clinical analogue has not been tested: an active, expensive imaging modality allocated per patient by expected information gain, inside a clinical-radiomic fusion model.
Core Claim
A clinical-radiomic model that adds perfusion and DTI features only for the patients a learned gain predictor selects can recover most of the full-cohort benefit of functional imaging at a fraction of the acquisitions. Concretely, the target is at least 80% of the C-index or AUC improvement while activating 50% or fewer of patients.
The gain predictor estimates, for each patient, how much the full model's loss would drop relative to the structural-plus-clinical model. The claim that would be new is that this predictor beats the obvious rule, "scan the patients whose current prediction is least confident." The reasoning is that confidence and information gain differ. A confidently wrong structural prediction, for example where treatment-like necrosis mimics tumor, is exactly where perfusion helps most, and a confidence threshold never selects it.
The motivating evidence that structural imaging degrades in exactly such tissue states comes from the cervical cancer cohort in Evaluation of Chemotherapy Response with Serum SCC-ag. There, MRI agreement with pathology fell from 95% to 60% after chemotherapy-induced necrosis and fibrosis, while a complementary signal (serum antigen) lifted the validation AUC from 0.734 to 0.950. That paper fuses for everyone. We ask whether the complementary signal is needed for everyone.
Where This Sits Relative to Prior Work
The clinical-radiomic literature in the file supplies the base pipeline and the case for fusion, but it always assumes uniform acquisition. Two papers fuse radiomic scores with clinical variables via regression or survival models: Li et al.'s PET glioma model reached 0.900 validation AUC for IDH genotype, and Leu et al. on NLST raised survival C-index from 54.81 to 67.06 by adding shape radiomics to clinical data. Ye et al. show that adding quantitative dynamic contrast parameters to structural radiomics improves IDH prediction. That is our premise that the active modality helps on average, but Ye et al. never ask for whom it helps. Leu et al. also constrain the design: high-order texture features hurt a Cox model relative to curated shape features. We therefore keep the per-modality feature sets small and regularized rather than adding capacity.
The wireless papers supply a design pattern, not mathematics, and we do not claim the optimization theory transfers. Three results matter. Huang et al. show that a jointly optimized hybrid converges in about half the iterations of alternating optimization. That motivates training the fusion model and the selection rule together rather than in sequence. Shu et al. report that activating half the elements gives a rate "close to MRR with negligible performance loss," which is the outcome shape we hope for. Mu et al. find that 3-bit phase control nearly matches continuous control, a reminder that coarse allocation can be enough.
The other direction the file suggests is placing physical reflecting surfaces inside MRI or microwave imaging hardware. We set it aside because nothing in the file provides data or a testbed for it.
Research Program
We define two feature sets. The passive set $x_i$ is structural MRI radiomics plus age, sex and resection status. The active set $z_i$ is perfusion and DTI radiomics, using the CaPTk features distributed with UPenn-GBM. Each patient $i$ has a base model prediction $f_P(x_i)$ and a full model prediction $f_A(x_i, z_i)$. Both are Elastic-Net Cox models for survival and logistic models for MGMT methylation, matching the regularized pipelines in the cited clinical papers.
The per-patient gain is $g_i = \ell(f_P(x_i), y_i) - \ell(f_A(x_i,z_i), y_i)$, where $\ell$ is the per-patient loss. For classification, $\ell$ is log-loss. For survival, $\ell$ is the patient's contribution to the negative Cox partial likelihood. Gains are computed out of fold so that $f_A$ is never scored on its own training patients.
A small gain predictor $h(x_i)$, which sees only passive features, is trained to rank patients by $g_i$. At deployment with activation budget $\rho$, the top $\rho$ fraction by $h$ receive the full model and the rest receive the base model. The joint variant retrains $f_A$ only on patients $h$ would select. It then iterates between refitting $h$ and $f_A$ under cross-fitting, so the full model specializes to the patients it will actually serve.
The analysis output is a curve of performance against the fraction activated. Alongside the curve, we characterize the selected patients, for example necrosis volume, enhancement pattern and resection extent, to test whether selection tracks interpretable radiologic ambiguity.
UPenn-GBM is the primary dataset because it is public via TCIA and contains both modality types plus clinical, survival, IDH (n=515) and MGMT (n=317) labels. All controlled comparisons use the complete-case subset that has both perfusion and DTI, so every policy is scored on identical patients. IDH is a secondary endpoint only: in de novo glioblastoma, IDH-mutant cases are rare, so the AUC will be noisy.
Baselines
The naive reference is the structural-plus-clinical model given to everyone, $f_P$, which corresponds to activation fraction zero. The ceiling is the full fusion model given to everyone, $f_A$ at activation fraction one, in the late-fusion style of Li et al. and Ye et al.. It is the strongest existing method, since it uses all information. Success is measured as the fraction of the gap between these two that a policy recovers at a given budget.
Neither number exists in the literature for this task: no reported result was found on Lacuna for survival, MGMT or IDH prediction on UPenn-GBM. The closest reported fusion figures are on different cohorts and modalities and serve only as context: 0.900 validation AUC for PET-based IDH prediction in Li et al., and C-index 67.06 for clinical plus shape radiomics on NLST in Leu et al.. We therefore establish both reference points ourselves under one protocol: identical patient splits (repeated stratified 5-fold, plus a held-out temporal split if acquisition dates permit), identical feature preprocessing, and an identical hyperparameter search budget per model.
The competing selection policies, all scored at matched activation fractions, are as follows. Random activation is a lower bound on any policy. Confidence thresholding activates patients whose base-model probability is nearest 0.5, or whose predicted survival distribution is widest. This is the strongest simple competitor and the comparison the paper rests on. An imputation model predicts $z_i$ from $x_i$ and uses the imputed values for everyone, testing whether the active modality carries information the passive one cannot reconstruct.
Two ablations isolate the key ingredient. The first replaces the learned $h$ with confidence while keeping the jointly trained $f_A$. The second keeps $h$ but uses an $f_A$ trained on all patients. Together they separate "choosing the right patients" from "specializing the full model to those patients." Beating a policy means a higher gap-recovered fraction at $\rho = 0.3$ and $\rho = 0.5$, with a paired bootstrap confidence interval over patients excluding zero.
Why not just…?
Why not just give everyone the full multimodal model? If acquisition were free, this would be correct, and it is our ceiling. But perfusion is missing for about one patient in five even in a research cohort (UPenn-GBM), so routine practice is already selective, only unprincipled. If the curve shows functional imaging's benefit spread evenly across patients, the honest conclusion is that selection is pointless and everyone should be scanned. That is itself a useful finding.
Why not just scan the patients the base model is least sure about? This is the real threat, and it may well work as well as a learned gain predictor. A small cohort leaves little signal for learning a second model, and a regularized $h$ may collapse to something like confidence anyway. The head-to-head at matched budget settles it. If confidence thresholding matches the learned predictor, we report that and recommend the simpler rule. The contribution then becomes the measured trade-off curve and the joint-training ablation rather than the selector.
Why not just impute the missing functional features from structural ones? Imputation spends nothing extra. If it closes most of the gap, the active modality is largely redundant with the passive one, which would undercut Ye et al.'s premise for this population. The imputation baseline directly measures how much of $z_i$ is recoverable from $x_i$. A large residual gap is the precondition for selective acquisition to matter at all.
Decisive Checks
The first check is the gap-recovery curve on complete-case UPenn-GBM for overall survival C-index and MGMT AUC. We expect the curve to be concave, with most gain from the first third of patients, if gains are concentrated. A linear curve would falsify the core premise.
The second is the matched-budget comparison of learned gain prediction against confidence thresholding. Following the reasoning above, we predict the learned selector's advantage appears mainly among confidently wrong base predictions. We will test this by stratifying recovered gain by base-model confidence.
The third is an external-validity probe. We fit the policy on patients with full acquisitions, then check whether the patients who naturally lack perfusion in UPenn-GBM resemble those the policy would have deprioritized. If they do, routine practice already approximates the policy. If they do not, the missingness pattern is informative and the complete-case results need that caveat.
Any claim of clinical readiness waits on a second, multi-site cohort, given the scanner-protocol sensitivity of radiomic features flagged throughout this literature.
References
- The University of Pennsylvania glioblastoma (UPenn-GBM) cohort: advanced MRI, clinical, genomics, & radiomics
- Coexisting Passive RIS and Active Relay Assisted NOMA Systems
- Three High-Rate Beamforming Methods for Active IRS-Aided Wireless Network
- Evaluation of Chemotherapy Response with Serum Squamous Cell Carcinoma Antigen Level in Cervical Cancer Patients: A Prospective Cohort Study
- A Non-invasive Radiomic Method Using 18F-FDG PET Predicts Isocitrate Dehydrogenase Genotype and Prognosis in Patients With Glioma
- Survival prediction of NLST lung cancer patients from radiomics and clinical data
- Integrating quantitative DCE-MRI parameters and radiomic features for improved IDH mutation prediction in gliomas
- Exploiting Intelligent Reflecting Surfaces in NOMA Networks: Joint Beamforming Optimization