Pre-Surgical Growth Kinetics with Scanner-Aware Uncertainty for Occult Nodal Metastasis

Many lung cancer patients have more than one CT scan of their nodule before surgery: a screening scan, an incidental finding, a surveillance follow-up. Yet models that predict hidden lymph-node spread ("occult nodal metastasis," cancer in nodes that look normal on imaging) use only the last preoperative scan. We propose to use the earlier scans as well, by estimating how fast each nodule grew. The obstacle is noise. Prior scans usually come from different machines, slice thicknesses and reconstruction settings, and over short intervals the volume error from those differences can be as large as the true growth. The proposal is to fit a small Bayesian growth model for each patient. In that model, the size of the measurement error is learned separately for each scanner protocol and shared across patients scanned with it. The resulting distribution over growth rate, not a single noisy number, is then passed into the nodal-metastasis classifier.

Two things would be new. The first is using pre-surgical growth trajectories for occult nodal metastasis at all. The second is treating multi-center scanner differences as a measurable error term inside the growth model, rather than as something to normalize away. A patient with only one scan simply inherits the population growth prior, so the method falls back to the static model and is never worse in coverage. The clinical question it serves is concrete: should this patient get a full lymph-node dissection or invasive nodal staging, or not? It also tells the surgeon when the answer is uncertain.

Core Claim

Growth rate before surgery carries information about occult nodal spread beyond what the final scan shows. That information is usable only when the measurement error from changing protocols is modeled explicitly. Concretely, we expect two results. The full model should improve leave-one-hospital-out discrimination and calibration over the strongest static fusion model. It should also beat a naive "volume doubling time as an extra feature" baseline specifically in patients whose prior scans are close in time or acquired on mismatched protocols. If the naive feature matches it everywhere, the Bayesian machinery is unnecessary for discrimination, and we will say so.

Prior Work and Where This Sits

The static detectors define the target to beat.

Predicting occult lymph node metastasis in solid-predominantly invasive lung adenocarcinoma across multiple centers using radiomics-deep learning fusion model fuses handcrafted radiomics with a pretrained ResNet-18 across six hospitals. It uses a momentum update to push toward scanner-invariant features. It reaches an average external AUC of 0.754, from one preoperative scan, with no uncertainty output.

Dual-layer spectral detector CT quantitative parameters and radiomics for predicting spread through air spaces of lung adenocarcinoma: a dual-center study predicts a related form of microscopic invasion, spread through air spaces (STAS). Its AUC falls from 0.910 in training to 0.848 externally, within a single vendor.

Both papers motivate this proposal: empirical invariance leaves residual site effects that a deterministic model can neither express nor bound. We differ from both by adding a time axis and by modeling the site effect rather than suppressing it.

Two papers show that change over time adds signal beyond one snapshot. AI-based non-invasive profiling of the tumor immune microenvironment using longitudinal CT radiomics predicts immunotherapy response in lung cancer computes a log volume change rate from serial CTs to predict immune phenotype (AUC 0.85). However, it does so after treatment starts, deterministically, and without multi-center harmonization. Integrated CT-derived radiomics for predicting ctDNA status and prognosis in resectable non-small cell lung cancer tracks circulating tumor DNA (tumor DNA shed into blood) over time, but keeps the imaging static. Its authors name scanner variation as unmodeled noise, which is the term we model.

The probabilistic machinery comes from Comparing Methods for Uncertainty Estimation of Paraganglioma Growth Predictions. That paper fits Bayesian growth parameters to serial MRI volumes and feeds segmentation uncertainty from five nnU-Net models into the fit. It also supplies the central constraint: its nominal 95% intervals covered the truth only 78% of the time, while a historical-error model covered 94% but with intervals wider than 200% of the predicted volume. We read the 78% as evidence that the measurement error was underestimated, likely because it did not account for protocol differences. Our design is a direct attempt to close that gap.

Personalized Radiotherapy Planning for Glioma Using Multimodal Bayesian Model Calibration shows that posterior inference over growth parameters can map invisible tumor infiltration, but on eight patients from a single time point. We deliberately use a much simpler growth law than its reaction-diffusion equation, because nodal spread from a lung nodule is not spatially contiguous with the primary tumor.

Research Program

Data. We will retrieve every available prior chest CT for patients in the six-hospital cohort, together with its acquisition metadata: vendor, slice thickness, reconstruction kernel and dose. We will also use the National Lung Screening Trial (NLST) serial low-dose screens for screen-detected adenocarcinomas with surgical staging. NLST offers regular intervals on a single protocol family, which serves as a contrast to routine-care irregularity.

Segmentation and its uncertainty. An nnU-Net ensemble segments each nodule's total and solid-component volume at every time point. Disagreement within the ensemble gives a per-scan segmentation uncertainty $\tau_{ij}$, following the paraganglioma study.

Protocol error. We estimate protocol error directly from data. Many CT exams store two reconstructions of the same acquisition, typically a sharp lung kernel and a soft-tissue kernel, and many centers archive thick and thin slices. Volume differences between reconstructions of the same acquisition isolate the error caused by protocol alone, with no biological change. These pairs inform priors on a bias $b_s$ and a spread $\sigma_s$ for each protocol group $s$.

Growth model. For patient $i$ with scans at times $t_{ij}$, we write

$$\log V_{ij} = a_i + g_i\, t_{ij} + b_{s(ij)} + \epsilon_{ij}, \quad \epsilon_{ij} \sim \mathcal{N}\!\left(0,\ \sigma_{s(ij)}^2 + \tau_{ij}^2\right),$$

where $V_{ij}$ is measured volume, $a_i$ is baseline log-volume, $g_i$ is the patient's exponential growth rate, and $s(ij)$ is the protocol of scan $j$. Growth rates are partially pooled, $g_i \sim \mathcal{N}(\mu_{c(i)}, \omega^2)$, with a separate mean $\mu_c$ for each center $c$, because referral patterns differ by center. Fitting uses standard Hamiltonian Monte Carlo. With a handful of parameters per patient, this is cheap, unlike the high-dimensional sampling criticized in the file.

Prediction. The nodal-metastasis probability integrates over the growth posterior:

$$P(N^{+}_i \mid \text{data}) = \int P(N^{+} \mid g, x_i)\, p(g_i \mid \text{scans}_i)\, dg,$$

where $x_i$ is the static fusion-model embedding of the final scan. The classifier $P(N^{+} \mid g, x)$ is trained with growth samples drawn from each training patient's posterior. Under this training, noisy growth estimates are automatically down-weighted, and single-scan patients reduce to the static model. The spread of the predicted probability supports a defer-to-invasive-staging rule.

Baselines

The naive reference point is clinical: tumor size and solid-component proportion in logistic regression, with no imaging model. This tells us how much any learned model adds.

The strongest existing method is the ResNet-18 plus radiomics fusion model with momentum update from Predicting occult lymph node metastasis in solid-predominantly invasive lung adenocarcinoma across multiple centers using radiomics-deep learning fusion model. It reports an average external AUC of 0.754, against 0.715 for radiomics-only and 0.676 for deep learning only. We will re-implement it on the identical patient set and splits. We will also report it under leave-one-hospital-out rotation, since a single fixed external split of two hospitals is fragile. The same fusion embedding $x_i$ feeds our model, so any gain is attributable to the time axis.

The direct competitor to the key ingredient is a naive kinetics feature. This is volume doubling time (or log volume change rate, as in AI-based non-invasive profiling of the tumor immune microenvironment using longitudinal CT radiomics predicts immunotherapy response in lung cancer), computed by least squares from the same scans and segmentations and concatenated to $x_i$. Imputation is used for single-scan patients.

Three ablations isolate our ingredients. The first uses a single global noise term instead of protocol-specific $\sigma_s$ and $b_s$. The second uses posterior-mean growth only, without integrating over uncertainty. The third drops the kernel-pair noise priors.

All methods receive the same segmentations, an equal hyperparameter search budget tuned only on training hospitals, and identical post-hoc temperature scaling. Calibration differences therefore cannot come from recalibration effort.

Beating the strongest baseline means two things. Pooled external AUC must improve over the fusion model, with a bootstrap confidence interval excluding zero. Calibration slope and expected calibration error must improve in at least four of six held-out hospitals. Beating the naive kinetics baseline means a significant AUC or net-benefit gain in the pre-specified subgroup of short-interval or protocol-mismatched patients. Matching it there would count as failure.

Why not just…?

Why not just add volume doubling time as a feature? This is the real threat, and it may match us when intervals are long and protocols constant, as in NLST. Our argument is quantitative rather than rhetorical. When interval growth is comparable to protocol-induced volume error, a least-squares growth rate is dominated by noise. The kernel-pair analysis will measure that error directly. The stratified comparison by interval length and protocol mismatch then settles whether the explicit error model matters. If it does not, the lasting contribution is the kinetics signal itself, and we will report it that way.

Why not just recalibrate or apply conformal prediction to the static model per site? Post-hoc calibration, as offered in PyHealth: A Deep Learning Toolkit for Healthcare Applications, can fix a static model's average coverage. It cannot add information, and it needs labeled data from each new hospital. The paraganglioma study shows the general pattern: uniform historical-style margins reach coverage (94%) only by becoming uninformative (more than 200% width). We include the temperature-scaled static model as a baseline. If it matches our calibration and our AUC gain is null, the proposal fails.

Why not just harmonize features or train for domain invariance? The momentum-update approach already does this and still ends at 0.754 external AUC. More importantly, invariance methods act on features from one scan. They cannot tell a nodule that grew from a nodule re-imaged with a thicker slice, which is exactly the confound in serial pre-surgical CTs. The global-noise ablation tests whether explicit protocol terms matter beyond what invariance achieves.

Proposed Experiments and Decisive Checks

The first check is a growth-forecast calibration test that is directly comparable to the paraganglioma result. For patients with three or more scans, we hold out the final preoperative volume, predict it from the earlier scans, and measure 95% interval coverage and width. We aim for coverage near nominal, with widths far below the historical-error regime. The global-noise ablation should reproduce under-coverage of the kind reported there (78%). If protocol-specific noise does not close most of that gap, the mechanism is wrong.

The second check is leave-one-hospital-out occult nodal metastasis prediction, reporting AUC, calibration slope, expected calibration error, and decision-curve net benefit at dissection-relevant thresholds. Results are stratified a priori by number of prior scans, interval length and protocol mismatch.

The third check is selective prediction. We defer the most uncertain cases to invasive nodal staging, then measure the false-negative rate among non-deferred patients as a function of deferral rate, compared against deferral based on the fusion model's predicted probability.

The fourth check uses the NLST serial screens as a near-ideal protocol setting. It tells us how much signal kinetics carries when noise is low, which separates "growth is uninformative" from "growth is informative but buried in protocol noise."