After an ischemic stroke, doctors scan the brain on arrival and try to predict which tissue will die. The ground truth they compare against is a follow-up MRI taken anywhere from 2 to 9 days later, and the stroke began some unrecorded-but-estimable time before the first scan. Current prediction models ignore both delays. They map the admission scan to the follow-up lesion as if time did not matter. Epidemiologists and tree-ring ecologists solved a similar problem long ago. When a driver acts with a variable delay, they fit a distributed lag non-linear model (DLNM): a smooth surface that describes how the outcome depends jointly on how strong the exposure was and how long ago it happened. We propose to fit exactly such a surface at the voxel level (a voxel is one 3D pixel of the scan). The exposure is perfusion severity, meaning how starved of blood a voxel looks on admission CT perfusion. The lag is the patient's elapsed ischemia time. The resulting "severity × duration" death surface is then used to correct the outputs of a standard deep segmentation network.
In the 2024 Ischemic Stroke Lesion Segmentation challenge (ISLES'24), the best deep model improved spatial overlap over a fixed clinical perfusion threshold (Dice 0.285 vs 0.163). Yet both had the same volume error, about 21.2 mL. No team made use of clinical timing data. We think these two facts are connected. A lesion measured on day 9 is systematically larger than the same lesion measured on day 2, and a patient scanned six hours after onset has lost more tissue than one scanned at one hour. A model blind to these intervals cannot get volume right, however good its spatial features are. The claim is narrow and checkable: most of the leftover volume error comes from unmodeled timing. A constrained lag-response surface can recover this with only a few hundred patients, where free-form conditioning of a deep network cannot.
Core Claim
Let $s_{iv}$ be the perfusion severity of voxel $v$ in patient $i$ (for example Tmax, the time for contrast to peak, or relative cerebral blood flow). Let $d_i$ be that patient's ischemia duration, meaning time from symptom onset to follow-up, interacted with reperfusion status where available. Let $c_i$ be the treating center. We model the probability that the voxel is infarcted on follow-up as
$$ \operatorname{logit} P(y_{iv}=1) = \alpha\, g_{iv} + \sum_{j,k} \theta_{jk}\, b_j(s_{iv})\, q_k(d_i) + u_{c_i}, $$
The terms are as follows. $g_{iv}$ is the logit output of a frozen nnU-Net, entered as an offset whose weight $\alpha$ is learned. $b_j$ and $q_k$ are low-order spline bases over severity and duration, and their product is the DLNM "cross-basis." $\theta_{jk}$ are the surface coefficients, constrained so that risk cannot decrease with duration. $u_{c_i}$ is a random intercept for each center.
The cross-basis comes directly from the scrub-typhus DLNM of Liao et al.. That study modeled weekly case counts as a joint non-linear function of meteorological intensity and lag, and found inverse-U intensity curves whose peaks depended on the lag. The center-level random effect comes from the multilevel DLNM of Zhao et al.. That work showed that a county's baseline climate shifts its lag window. We use the same device to absorb differences in scanner and protocol between hospitals.
Two things are new. First, the "lag" axis here is not a series of past observations. It is between-patient variation in a known interval, which lets one follow-up scan per patient identify the time dimension of the surface. Second, the surface is not a standalone statistical model. It acts as a structured correction on a deep network. The surface is also a clinically interpretable object in its own right: an effective perfusion threshold that loosens as ischemia lengthens, replacing the single fixed rCBF < 30% cutoff.
The motivating analogy is the tree-ring result of Zhang et al.. That study showed that vegetation models which omit multi-year lag effects overstate a forest's immediate sensitivity to climate. We expect the stroke version of the same error: a timing-blind model folds duration into its apparent sensitivity to severity. This should show up as volume errors that are biased and correlated with the follow-up interval.
Research Program
The work proceeds in three stages on the ISLES'24 dataset. That dataset pairs admission non-contrast CT, CT angiography and perfusion maps with follow-up MRI lesion masks and clinical tables for 245 patients, and its lesion masks were refined to inter-rater Dice 0.86–0.90.
The first stage is purely diagnostic and gates everything after it. We reproduce a strong nnU-Net baseline in patient-level cross-validation and regress its per-patient signed volume error on onset-to-imaging time, imaging-to-follow-up time and reperfusion status. If these intervals explain essentially none of the error, the core hypothesis is wrong and we stop. We would then report that negative result, which is itself useful given that no ISLES'24 team tested it.
The second stage fits the surface. We start with severity alone and then add the full cross-basis. Spline degree is chosen by grouped cross-validation, and monotonicity in duration is enforced by constrained coefficients. We then add the center random effect and compare.
The third stage is the analysis that matters most. We examine where the surface helps: stratified by lesion size, with particular attention to the <5 mL regime where winning ISLES'24 models reached only 14.94% lesion-wise F1, and stratified by follow-up interval. We also check whether the implied time-varying threshold is physiologically plausible.
Clinical-radiomics studies of later stroke complications motivate this design and bound what we claim. The malignant edema model of Zhang et al. raised validation AUC from 0.818 to 0.916 by adding clinical factors in a simple logistic model. The hemorrhagic transformation study of Huang et al. reached AUC 0.90 from admission CT texture. Together they show that tabular information does help once it enters through a low-dimensional statistical layer rather than end-to-end. Both, however, treat their outcomes as static binary labels at one checkpoint. Neither has public data here, so we mention them only as a possible later transfer of the same surface idea to time-to-complication.
Baselines
The naive reference is the clinical perfusion threshold, rCBF < 30%, applied to the admission maps. ISLES'24 reports Dice 0.163 and AVD (absolute volume difference between predicted and true lesion volume) of about 21.2 mL for it on the 98-case hidden test set. We re-run it in our cross-validation folds so that every method is scored on the same patients.
The strongest existing method is the winning ISLES'24 approach, a 3D large residual-encoder nnU-Net. It reports Dice 0.285, AVD about 21.2 mL, and 14.94% lesion-wise F1 on lesions under 5 mL (ISLES'24). We retrain nnU-Net with its standard self-configuration on the same folds and the same input channels. It gets the same GPU budget as our pipeline's nnU-Net, which is in fact the identical frozen network, so the only difference is the surface layer. If the organizers' hidden test set is still accepting submissions, we report there too, so the result can be compared directly with the published 0.285.
Beating nnU-Net means a clear AVD reduction on paired per-patient comparisons with Dice no worse, rather than a Dice gain alone. The hypothesis concerns volume and timing, and the tie at about 21.2 mL is the gap we target.
Two ablations isolate the key ingredient. The first is a severity-only correction, the same model with the duration basis removed. It tests whether any gain comes from recalibrating nnU-Net rather than from time. The second is a single-lag version that replaces the smooth duration basis with one linear duration term. It tests whether the distributed, non-linear shape matters, echoing the finding in Liao et al. that a single driver can flip sign across lag windows.
Why not just…?
Why not just feed the timing variables into nnU-Net as extra inputs or feature-wise conditioning? This is the serious competitor, and we run it. Onset-to-imaging and imaging-to-follow-up times enter as constant input channels, and separately through feature-wise affine modulation of the decoder, under the same tuning budget. The evidence that it may fall short is indirect: ISLES'24 reports that no deep team turned tabular clinical data into better segmentation. With a couple of hundred patients, an unconstrained conditioning pathway can ignore the timing variables or overfit them. If conditioning matches the surface on AVD, we would conclude that the parameterization is not the contribution. The remaining value would then be the interpretable time-varying threshold, and we would say so.
Why not just use a time-dependent perfusion threshold? A lookup table of rCBF cutoffs by duration bin is the lowest-capacity version of our idea, and we include it. It lacks the deep network's spatial context, so we expect it to improve on the fixed threshold's AVD but trail nnU-Net on Dice. It is also the fairest test of whether the smooth, monotone surface adds anything over binning.
Why not just post-hoc rescale nnU-Net's predicted volume using the follow-up interval? Scaling total volume by a fitted function of time could close much of the AVD gap without touching voxels. If our stage-one diagnostic shows timing explains volume error, this rescaler becomes a mandatory baseline. The surface must beat it on Dice and small-lesion F1, because only a voxel-level model can decide where the extra tissue dies, namely at moderately hypoperfused voxels that cross threshold with time. If the surface only matches the rescaler on AVD and does not improve localization, the voxel-level claim fails.
Decisive Checks
The first go/no-go test is the stage-one regression of nnU-Net volume error on timing intervals. The main result is a paired, per-patient comparison of AVD, Dice and lesion-wise F1 between the frozen nnU-Net and nnU-Net plus surface, across the same folds, with confidence intervals from bootstrapping over patients. Stratifying by follow-up interval should show the largest gains in patients scanned late or followed up late. Absence of that pattern would mean any improvement is recalibration, not timing.
A leave-one-center-out split tests whether the center random effect earns its place, following the logic of the county-level heterogeneity in Zhao et al. Finally, the fitted surface should imply a severity cutoff that relaxes with longer duration and is roughly flat after reperfusion. A surface that contradicts basic ischemia physiology would signal confounding, for example sicker patients arriving later, and we would test this by adjusting for admission NIHSS (the standard stroke severity score).
References
- ISLES'24: Improving final infarct prediction in ischemic stroke using multimodal imaging and clinical data
- The Temporal Lagged Relationship Between Meteorological Factors and Scrub Typhus With the Distributed Lag Non-linear Model in Rural Southwest China
- The temporal lagged association between meteorological factors and malaria in 30 counties in south-west China: a multilevel distributed lag non-linear analysis
- Converging Climate Sensitivities of European Forests Between Observed Radial Tree Growth and Vegetation Models
- ISLES'24 -- A Real-World Longitudinal Multimodal Stroke Dataset
- A CT-based machine learning model for using clinical-radiomics to predict malignant cerebral edema after stroke: a two-center study
- The value of CT-based radiomics in predicting hemorrhagic transformation in acute ischemic stroke patients without recanalization therapy