Scanner Noise Lives in the Variance: Count-Level Sparse Radiomics for Unseen Hospitals

Radiomics models predict tumor grade or patient outcome from hand-crafted image measurements. They routinely lose accuracy when moved to a new hospital. A seven-center pancreatic cyst study saw accuracy fall from 71.0% under cross-validation to 61.6% on unseen centers, and found 3D texture features too unstable across slice thicknesses to use at all (Multi-center evaluation of radiomics and deep learning to stratify malignancy risk of IPMNs). The usual pipeline goes wrong at a specific place. Texture features are summaries of count tables: how often two gray levels sit next to each other, how many runs of a given length appear, how many voxels fall in each intensity bin. A scanner change mostly alters how noisy those counts are, not only their average. The pipeline collapses the counts to scalar summaries, treats them as Gaussian, may mean-match them across sites, and then runs LASSO, a sparse regression that sets most coefficients exactly to zero. By the time selection happens, the information about which counts are noisy at which scanner is gone.

We propose to select features at the count level using a negative binomial model, a count distribution with an extra "dispersion" parameter that lets the variance grow faster than the mean. The dispersion of each texture-table cell is written as a sparse function of acquisition metadata read from the image header: slice thickness, field strength, vendor, reconstruction kernel. A second sparse term picks the cells whose mean differs between diagnostic classes. The classifier then down-weights each cell by its predicted dispersion at the deployment scanner. Because that dispersion comes from header metadata alone, the model adapts to a new hospital without target labels and without a batch of target images.

Core Claim

Scanner-induced instability in radiomics is mostly heterogeneous over-dispersion of the underlying texture counts. Our claim is that modeling it explicitly lets three things happen. Features can be selected for class signal, separately from acquisition noise. The model can be specialized to an unseen site from metadata alone. And the 3D texture features that current practice discards become usable again.

Each of the three directions plays a distinct role here. Over-dispersion modeling supplies the mechanism. Sparse selection is the estimation constraint. Multi-center imaging shift is the regime and the test. The claim would be false in two cases: if a Gaussian location-and-scale model of the same structure on log-counts does equally well, or if nothing improves on unseen centers beyond ComBat harmonization followed by LASSO.

Model

For patient $i$ and texture-table cell $j$, let $y_{ij}$ be the count, $c_i$ the class label, $s_i$ a vector of acquisition descriptors, and $n_i$ the region-of-interest size, which enters as an exposure offset. We model

$$ y_{ij} \sim \mathrm{NB}(\mu_{ij}, \phi_{ij}), \quad \log \mu_{ij} = \log n_i + \alpha_j + \beta_j c_i + \delta_j^\top s_i, \quad \log \phi_{ij} = \gamma_{0j} + \gamma_j^\top s_i, $$

with variance $\mu_{ij} + \phi_{ij}\mu_{ij}^2$. Here $\beta_j$ is the class effect on cell $j$, $\delta_j$ captures scanner shifts in the mean, and $\gamma_j$ captures scanner-dependent dispersion. We fit by negative log-likelihood with separate $L_1$ penalties on $\{\beta_j\}$, $\{\delta_j\}$ and $\{\gamma_j\}$.

This directly reuses the double-penalized heterogeneous negative binomial estimator of Heterogeneous Overdispersed Count Data Regressions via Double-Penalized Estimations. That paper proved oracle inequalities and showed lower error than fixed-dispersion regression, but only on tabular health-survey counts with no site structure. We move the dispersion covariates from patient demographics to scanner physics.

At test time, a new scan's header gives $\hat\phi_{ij}$ and the $\delta$-adjusted means for both classes. The scan is then scored with a discriminant built from the negative binomial likelihood over the selected cells. This follows NBLDA: negative binomial linear discriminant analysis for RNA-Seq data, which down-weights high-dispersion features and beat Poisson discriminant analysis, support vector machines and nearest neighbors on RNA-seq data with median dispersions of 11.0 to 40.1. NBLDA neither selects features sparsely nor lets dispersion depend on the site. Both are what we add. Those dispersion values come from RNA-seq, not imaging; whether texture counts are comparably over-dispersed is checked in the first stage of the program.

Research Program

The work has four stages.

The first stage measures whether the premise holds. On AMOS, a 600-scan multi-vendor CT and MRI benchmark with 15 annotated abdominal organs (AMOS), we extract gray-level co-occurrence, run-length and histogram count tables within organ masks. We then ask how much of the cross-scanner variation for the same organ lives in dispersion rather than mean. If mean shifts dominate, the method reduces to a count-based ComBat and we will say so.

The second stage builds the estimator and the metadata-driven discriminant described above. Gray levels are quantized coarsely so that cell counts are not dominated by zeros.

The third stage covers selection. Neighboring table cells are strongly correlated, which violates the independence assumption that NBLDA makes. Plain $L_1$ would then pick cells arbitrarily within a correlated block. We therefore add a group penalty over each scalar feature's cells and report selection stability across folds.

The fourth stage tests the claim about 3D features. We include 3D tables, put slice thickness into $s_i$, and ask whether they now add signal across centers.

Segmentation masks come from the expert regions of interest where available. Elsewhere we use the PanSegNet pancreas model, which reports T2W MRI Dice of 86.3% (Large-Scale Multi-Center CT and MRI Segmentation of Pancreas with Deep Learning). We also test sensitivity to mask choice.

Baselines

The naive reference is LASSO logistic regression on standard scalar radiomic features with no harmonization. It is the workflow used in the three-center intraductal papillary mucinous neoplasm (IPMN) nomogram study, where a 9-feature signature reached validation AUC 0.811 and 0.822 alone, and 0.884 and 0.876 with clinical markers (Radiomic nomogram based on MRI to predict grade of branching type intraductal papillary mucinous neoplasms of the pancreas: a multicenter study). The same workflow underlies the colorectal CT model with external AUC 0.814 (Construction of a prognostic survival model for colorectal cancer patients using CT image texture analysis: a prospective cohort study).

The strongest established competitor is ComBat followed by LASSO. ComBat is an empirical Bayes method that removes per-site shifts in mean and scale while borrowing strength across features (Evaluation of ComBat harmonization for reducing across-tracer differences in regional amyloid PET analyses). It is the current default for multi-site imaging, and it already models a per-site scale. We give it its best case: at each held-out center it receives all of that center's unlabeled scans to estimate batch parameters, an advantage our method does not use.

We also run the reported pipeline of the seven-center IPMN study, which fuses 2D radiomics, a Random Forest and DenseNet121. That paper reports external AUC 69.2% and accuracy 61.6% under leave-center-out testing (Multi-center evaluation of radiomics and deep learning to stratify malignancy risk of IPMNs).

The closest robust-selection method is Distributionally Robust Feature Selection (Distributionally Robust Feature Selection). It relaxes subset selection by injecting Gaussian noise into features and minimizes worst-group loss, and it was shown only on tabular census data. We run it on log-transformed scalar features with training centers as groups. No reported result was found on Lacuna for it on imaging data.

Three ablations isolate the key ingredient. The Gaussian twin keeps the identical mean and log-variance structure on log-counts with a Gaussian likelihood, isolating the count likelihood. Constant-dispersion negative binomial sets $\gamma_j = 0$, isolating heterogeneous dispersion. The scalar-feature version applies our model to summarized features, isolating the count level. Poisson is included as the zero-dispersion limit.

All methods use the same masks and the same leave-one-center-out splits. Each gets the same nested-cross-validation tuning budget over penalty strengths and the same feature-count range. Beating the baselines means a higher pooled leave-center-out AUC than ComBat+LASSO under a paired DeLong test, a higher worst-center AUC, and a higher external AUC than the 69.2% fusion number without using any deep features.

Why not just…?

Why not just ComBat followed by LASSO? This is the real threat. ComBat already removes a per-site scale, so it may capture most of what our dispersion term does. Our method differs in three ways. It does not need target-site images. It ties dispersion to physical descriptors rather than site identity, so it can extrapolate to an unseen slice thickness. And it lets selection see the noise structure. If ComBat with access to the target scans matches us on pooled and worst-center AUC, the honest conclusion is that our advantage is deployment convenience, not accuracy. We will report it that way.

Why not just filter features by test-retest or cross-site reliability and then run LASSO? Reliability filtering discards a feature as a whole. Our model keeps a feature whose class signal survives once its scanner-dependent noise is down-weighted. The scalar-versus-count ablation and the 3D-reinstatement test will show whether that finer treatment matters, compared with filtering at the same feature budget.

Why not just a Gaussian heteroscedastic model on log-features? For large counts this may be nearly equivalent, because the negative binomial is then approximately log-normal. The Gaussian twin ablation tests this directly. We expect the negative binomial to help mainly in sparse high-order cells and small regions of interest, where log transforms of near-zero counts are unstable, as argued for small counts in Dimension Reduction for High-dimensional Small Counts with KL Divergence. If the twin matches us overall, we will report the regime-specific result instead.

Why not just deep domain adaptation from pixels? In the eight-method dermoscopy benchmark, adversarial methods improved all ten target sets, but their gains varied widely under low disease prevalence (Mitigating the influence of domain shift in skin lesion classification). With 359 scans spread across seven centers, data are thinner than in that benchmark. We include a domain-adversarial network (DANN) trained with target-center images as a comparison point rather than assuming it fails.

Proposed Experiments

The primary test is leave-one-center-out on the seven-center, 359-scan T2W IPMN cohort. It depends on obtaining access to that cohort, and the three-center BD-IPMN dataset is the fallback. With roughly 50 scans per held-out center, per-center AUC intervals will be wide. We therefore prespecify pooled AUC and worst-center AUC as primary endpoints and report bootstrap intervals.

The mechanism check runs on AMOS. Fit dispersion on all but one scanner. Predict cell-level variance for the same organ on the held-out scanner. Score that prediction by held-out log-likelihood against the Gaussian twin and against constant dispersion. If held-out dispersion is not predictable from metadata, the zero-shot claim fails regardless of downstream AUC.

The decisive downstream check has two parts. First, do 3D tables combined with slice-thickness-dependent dispersion beat 2D-only features across centers, reversing the 3D finding of the seven-center study? Second, does the selected cell set stay stable across folds, measured by Jaccard overlap, more than LASSO's? If both hold and ComBat+LASSO with target scans does not match us, the claim stands.