Forecasting Tumor Growth From Patients Who Stayed

Models that forecast a patient's future scan are trained on the scans that happen to exist. In aggressive diseases, those scans are not a fair sample. Patients whose tumors grow fastest die or become too sick to image, so late-time training examples come mostly from slow progressors. A forecaster trained this way learns that tumors grow slowly. It then tells a newly diagnosed patient, whose fate is unknown, that their tumor will grow slowly too.

We propose a way to measure this error, which current benchmarks cannot see, and to correct it. The correction treats "was this patient still being scanned at time $t$" as something the model predicts from the images themselves, and reweights the forecasting loss accordingly. The core bet is that a patient's own imaging history explains much of why they leave follow-up. If it does, a dropout model that reads the images can make a standard statistical correction valid where the usual corrections, built on tabular covariates, are not.

Core Claim

Continuous-time imaging forecasters are evaluated on held-out scans from the same filtered population they were trained on. CRONOS is one such model: it learns a velocity field over 3D MRI with flow matching and integrates it to arbitrary future times. Because training and test data share the same filter, the resulting bias never shows up in the reported error.

The clinical evidence that this filter is severe comes from a separate literature. Parkinson et al. followed 41 high-grade glioma patients for 12 months. 56% died or deteriorated, and the surviving 39% looked stable on cognitive measures, a "stability paradox" produced entirely by who remained. CRONOS reports its glioma results on Lumiere, a sparse, irregular high-grade glioma MRI series, and conditions on nothing about survival.

We claim three things, each testable:

  1. Survivor-trained forecasters systematically underestimate growth for patients at high risk of leaving follow-up.
  2. This bias can be exposed with a controlled-attrition protocol in which the dropout mechanism is known.
  3. Inverse-probability weighting with a dropout model that reads the latent image trajectory removes most of the bias. Weighting on tabular covariates alone does not.

Pairwise comparison benchmarks such as MS-CXR-T are largely immune to this problem. The act of having a pair is shared by training and test queries, so selection is the same on both sides. The damage is specific to forecasting, where the query patient has not yet been filtered. Saying where attrition matters, and where it does not, is part of the contribution.

Research Program

Stating the target. A forecast for a patient who may die needs a stated target, because a dead patient has no future tumor volume to predict. We adopt the survival-conditional forecast: the patient's scan at time $t$ given that they are alive and still eligible for imaging at $t$. We report it alongside a predicted exit hazard, so a clinician sees both "if still here, the tumor will look like this" and "probability of still being here."

Even this more modest target is biased by naive training. Among living patients, the ones who get scanned at $t$ are selected by visit schedules and by clinical suspicion. DeepJoint's "clinical presence" work showed that this kind of selection is informative in ICU laboratory data (DeepJoint).

The model. We attach an exit-and-visit head $h_\phi$ to the forecaster's latent state $z_{ij}$ at each past scan $j$ of patient $i$. The head predicts the probability of leaving follow-up before the next scan. We then train the flow-matching objective with stabilized inverse-probability weights:

$$\mathcal{L}(\theta)=\sum_i\sum_k \hat w_{ik}\,\lVert v_\theta(x_{ik},t_{ik})-u_{ik}\rVert^2,\qquad \hat w_{ik}=\prod_{j<k}\frac{1-\bar h_j}{1-h_\phi(z_{ij})}.$$

Here $v_\theta$ is the forecaster's predicted velocity at scan state $x_{ik}$ and time $t_{ik}$, and $u_{ik}$ is the flow-matching target velocity. $\bar h_j$ is the marginal exit rate at step $j$; it keeps the weights stable. Weights are clipped at a fixed percentile.

This weighting is valid if dropout is "missing at random given the observed imaging history." In plain terms, whatever drives leaving follow-up must be visible in scans already taken. This assumption is far more plausible than the same assumption given only age, grade and treatment, because tumor growth visible on the last scan is exactly what predicts death.

Dropout driven by growth that happens between the last scan and death violates the assumption. We probe that case with a sensitivity analysis rather than hide it. The analysis shifts the unobserved dropouts' growth rates by a range of delta values and reports how far conclusions move.

The evaluation protocol. We build a controlled-attrition version of Lumiere. We take patients with the longest series and measure each one's true growth rate from their full series. We then truncate their series with a probability that rises with that growth rate, using a few mechanism strengths, including one calibrated to Parkinson's roughly 56% one-year exit.

Models are trained on the truncated data and scored on the withheld later scans. This gives ground truth for exactly the patients a real cohort loses. The full series are themselves survivor-selected, so the protocol measures the incremental bias and its correction, not absolute natural history. We state this plainly.

The endpoint. Because image metrics and clinical endpoints diverge, we segment every forecast image and score whole-tumor volume. Ranjbar et al. report nnU-Net whole-tumor Dice of 0.80 on post-treatment scans and necrotic-core Dice of only 0.35. We therefore use whole tumor as the endpoint and treat necrosis as unreliable.

The real-data check. On chest radiographs, we link MIMIC-CXR sequences to MIMIC-III admission and mortality records and forecast next-study progression labels at a fixed horizon. There is no ground truth for the patients who leave. We instead test a prediction of the theory: after correction, forecast severity should rise most in the high-exit-risk group. We also check whether uncorrected forecasts are miscalibrated in the direction of the stability paradox, using patients who returned later than scheduled as a partially observed check.

Baselines

Last Context Image. The "do nothing" reference predicts that the future scan equals the most recent one. It is not a straw man. CRONOS reports that ConvLSTM and SimVP failed to beat this heuristic, and that CRONOS itself reached 7.55 normalized root mean squared error (NRMSE) against the heuristic's 8.38 on Lumiere. It is run on the identical truncated splits. It encodes zero growth, so any method that cannot beat it on volume error for high-risk patients has failed.

Uncorrected CRONOS. This is the strongest existing forecaster. It is the only continuous-time model in the file evaluated on longitudinal glioma MRI, and 7.55 NRMSE on Lumiere is the number to reproduce on the untruncated split before anything else. All weighted variants share its architecture, training compute, number of epochs and hyperparameter search budget, so the only difference is the loss weights. Beating it means lower growth-rate bias and lower volume error in the top exit-risk tertile under injected attrition, without worse NRMSE overall.

CRONOS with conventional inverse-probability-of-censoring weighting. This baseline weights by a Cox model fit to tabular covariates only: age, grade, extent of resection and treatment. It is the obvious standard fix. Comparing it with our image-latent weights is the ablation that isolates the key ingredient: whether the dropout model must read the images.

Joint training without weighting. A second ablation trains the exit head jointly, in the style of DeepJoint, but does not use it to weight the loss. This separates the benefit of shared multi-task representation from the benefit of reweighting. DeepJoint reports a C-index of 0.878 at a 1-day horizon on MIMIC-III ICU labs. That number is not comparable to our setting and serves only as a reference for how well a jointly trained exit head can rank risk.

Chest radiograph forecasting. For this task, the comparator is a CoCa-CXR-style pair model repurposed as a forecaster (CoCa-CXR), which reports 65.0% progression accuracy on MS-CXR-T. That figure is for pairwise comparison, not forecasting. No reported forecasting result was found on Lumiere or MIMIC-CXR for any attrition-aware model.

Why not just…?

Why not just use conventional inverse-probability weighting on clinical covariates? This is the real threat. If tumor grade and treatment already explain dropout, image-based weights add nothing. Wang, Ying and Xu show that conditional-independence corrections fail when latent health drives selection. Their evidence comes from left truncation in an aging cohort, not glioma imaging, so it motivates our hypothesis rather than proving it. The tabular-weighting baseline settles the question directly. If it closes the bias as well as our weights under the calibrated mechanism, we will report that. The contribution then becomes the protocol and the finding that cheap weighting suffices.

Why not just condition the forecaster on survival or time to death? Future survival is unknown at deployment, so conditioning on it leaks the outcome. Conditioning on baseline covariates is what the tabular-weighting arm already does. Predicting the exit hazard as an output, which we do, is the deployable version of this idea.

Why not just accept the survivor-conditional forecast as the right target and skip correction? We do accept it as the target. Naive training still does not estimate it, because among living patients, who gets scanned at $t$ depends on suspicion and schedules. The injected-attrition experiment includes a variant with visit-driven missingness only and no deaths. Its purpose is to show whether this gap is real or negligible.

Why not use a proximal bridge estimator as in the PQB work? Its key assumption, that the chosen proxies capture the hidden confounding, cannot be tested from data. It was also built for left truncation in low-dimensional records. We keep the simpler assumption, missing at random given the image history, and bound its failure with sensitivity analysis rather than add a second untestable layer.

Proposed Experiments

The decisive experiment is controlled attrition on Lumiere, repeated at several mechanism strengths and seeds. The primary metric is forecast growth-rate bias: predicted minus true whole-tumor volume change, from segmented forecasts, reported by exit-risk tertile. Secondary metrics are NRMSE, and discrimination and calibration of the exit head.

The hypothesis fails if uncorrected CRONOS shows no growth underestimation in the high-risk tertile as attrition strengthens. It also fails if image-latent weighting does no better than tabular weighting. Either outcome is informative and would be reported.

Two further checks guard against misleading wins. A null-mechanism control, with random truncation, should show no benefit from weighting. Any gain there would indicate that the weighting helps through regularization rather than bias correction.

The MIMIC-CXR linkage then tests whether the correction moves real forecasts in the predicted direction, concentrated in high-exit-risk patients. The delta-adjustment sensitivity analysis reports how much growth between the last scan and death would be needed to overturn the conclusions.