When a blocked brain artery is reopened after a stroke, some tissue still dies. Part of it was already dead before treatment. Part of it starved during the hours of delay. A further part dies even though the large vessel was reopened, because the tiny downstream vessels stay shut. This last failure is called no-reflow. Today's best infarct-prediction systems blur these three causes together. They read the admission scans and output one static map of the final infarct. A doctor cannot ask such a map "how much more tissue would we lose if reperfusion came two hours later?", and it cannot separate time-driven loss from no-reflow-like loss. Continuous disease-progression models have spent a decade solving a closely related problem in chronic disease: placing patients on a continuous timeline and forbidding biologically impossible recovery. That work has never been applied to the minutes-to-hours window of acute ischemia.
We propose a voxel-wise "tissue clock." For each brain location, the network predicts three things: the chance the tissue is already dead, a rate at which surviving-but-starved tissue (the penumbra) dies per hour of untreated ischemia, and a residual chance of death that does not depend on time, which absorbs failure despite reperfusion. These combine through a closed-form survival expression. Elapsed ischemic time, taken from each patient's clinical record, enters explicitly. Because patients arrive after very different delays, the cohort as a whole traces out the time axis, even though each patient contributes only one follow-up scan. The model therefore predicts final infarcts that are guaranteed to grow, never shrink, as delay increases. It also yields a map of tissue death not explained by elapsed time, which is a testable candidate signature of no-reflow.
Core Claim
On the ISLES'24 longitudinal stroke benchmark, a structured, time-conditioned survival head should match the challenge-winning segmentation network on spatial accuracy. It should also beat that network on two properties a static network does not guarantee: monotone, calibrated counterfactual response to treatment delay, and stratified accuracy across short- and long-delay patients. If simply feeding time variables into the winning network achieves the same thing, the structural part of our proposal is unnecessary. The experiments are designed to tell those outcomes apart.
Background and Positioning
The benchmark is ISLES'24 (ISLES 2024 dataset; ISLES'24 challenge report). It pairs acute CT imaging with follow-up diffusion MRI at 2–9 days for 245 cases. The acute imaging includes non-contrast CT, CT angiography, 4D CT perfusion (a movie of contrast flowing through the brain) and its derived summary maps of blood flow, blood volume, and transit times. The cases also carry tabular metadata such as onset-to-door time and NIHSS (a stroke severity score). The challenge report motivates our work. Its winning 3D nnU-Net, a self-configuring medical segmentation network, reached a Dice overlap of 0.285 against 0.163 for the clinical rule "relative blood flow below 30%." Both scored about 21.2 mL absolute volume error. Lesion-wise F1 on small infarcts was 14.94%. The report also notes that top entries did not exploit clinical timing variables and collapsed the raw perfusion movie into summary maps. We take the task and the metrics from this benchmark. We change the formulation from a two-point image translation to a time-indexed survival prediction.
Clinical work motivates making time explicit. The stroke treatment review (Progress in the Treatment of Acute Ischemic Stroke) reports that each hour of delay to perfusion raises treatment failure by about 7%. It also reports that salvageable penumbra persists in many patients at 6 and even 18 hours, so a fixed time window and a fixed perfusion threshold are both crude. The mechanistic literature explains why reopening the artery may not be enough. Pericytes, the contractile cells wrapping capillaries, stay clamped after recanalization (Xiaoxuming Decoction review). Inflamed endothelium traps leukocytes (Dl-3-n-butylphthalide study). Leaky vessels cause edema that compresses capillaries from outside (Nuciferine study). Coronary no-reflow in PCI shows the same pattern (Huangshan ACS cohort). These papers motivate a time-independent residual death term. They also limit what we can claim: none of them provides imaging ground truth for no-reflow, so our residual map is a hypothesis, not a diagnosis.
The method comes from progression modeling. DFPL (Forecasting Irreversible Disease via Progression Learning) factorizes future disease probability as current status plus progression of the remainder, which guarantees no recovery. It was tested only on childhood retinal atrophy over school years, and it treats all pathology as uniformly irreversible. Surreal-GAN (Yang et al.) and the monotone networks of Wang et al. impose monotonicity along latent disease axes for the same reason. The survey by Sow and Oxtoby describes how latent-time models align patients seen at different disease stages onto one timeline. We borrow that idea, but our time axis is observed (minutes since onset) rather than latent. BrainODE (Park et al.) and the GRU trajectory model of Jung et al. model continuous trajectories from many visits per subject over years. ISLES'24 offers one follow-up per patient, so a neural ODE would be fitting a curve through two points. We deliberately use a closed-form hazard instead.
Research Program
A shared 3D encoder–decoder reads the acute imaging and outputs three non-negative maps per voxel $v$:
- $c_v \in [0,1]$, the probability the voxel is already irreversible core;
- $\lambda_v \ge 0$, a penumbral death rate per hour of ischemia;
- $r_v \in [0,1]$, a time-independent probability of death despite reperfusion.
The predicted probability that voxel $v$ is infarcted at follow-up, given the patient's ischemic duration $\tau$ in hours, is
$$p_v(\tau) = 1 - (1 - c_v)\, e^{-\lambda_v \tau}\, (1 - r_v).$$
This expression is non-decreasing in $\tau$ by construction. It reduces to DFPL's factorization when $r_v = 0$ and the exponential term is replaced by a learned progression probability. Unlike DFPL, irreversibility applies only to the core term, while the penumbral term depends on how long reperfusion was delayed.
The ischemic duration $\tau$ is onset-to-treatment time when the metadata provide treatment timings. For patients without them, it is onset-to-acute-imaging plus a learned offset for that missingness pattern. We will report results separately for the subgroup with complete timing.
Training uses voxel-wise binary cross-entropy plus a Dice term on $p_v(\tau_{\text{patient}})$. This is identical to the baseline loss, so differences come from the structure, not from the objective.
Identifying $\lambda$ separately from $r$ relies on spread in $\tau$ across patients: time-dependent loss must be explained by $\lambda$, time-independent loss by $r$. We will check this directly with a synthetic study. In it, we simulate infarcts from known $c$, $\lambda$, $r$ maps on real ISLES'24 perfusion inputs with real $\tau$ distributions, then measure how well the true maps are recovered.
A secondary arm swaps the summary perfusion maps for the raw 4D perfusion curves, compressed per voxel by a small temporal encoder. The question is whether residual death $r_v$ is better predicted from microvascular curve shape than from summary maps. Pericyte clamping and capillary plugging should alter how quickly contrast leaves the tissue, which deconvolved summary maps may not capture.
Baselines
The naive reference is the clinical threshold "relative blood flow below 30%," at the challenge-reported Dice of 0.163 and about 21.2 mL volume error. We re-run it on our folds.
The strongest existing method is the ISLES'24 winner, a 3D nnU-Net with a large residual encoder, at Dice 0.285, about 21.2 mL volume error, and small-lesion F1 of 14.94% on the hidden test set (ISLES'24 report). It is the right competitor because it is the best spatial predictor known on exactly this data. Our model reuses its architecture and nnU-Net preprocessing, with only the output head replaced. That makes the comparison as fair as possible: same five-fold cross-validation on the training cases, same epochs and augmentation, and an equal hyperparameter search budget. If the hidden test server accepts submissions, we report there as well. Beating it means matching its Dice within cross-validation noise while winning on delay-stratified Dice and on counterfactual consistency. A Dice gain alone, without those, would not support the core claim.
The key ablation is nnU-Net with $\tau$ and other clinical scalars broadcast as extra input channels or injected through feature-wise modulation, with an unconstrained sigmoid output. This isolates whether the gain comes from structured time dependence or merely from giving the network the clock.
A second ablation removes $r_v$ (pure core plus time-driven loss). It tests whether a time-independent residual term is needed at all.
A third ablation is a time-dependent threshold rule. Here the blood-flow and Tmax cutoffs are fit as functions of $\tau$ on the training folds. It is the cheapest way to make the clinical rule time-aware.
Why not just…?
Why not just add onset time as an input channel to nnU-Net? This is the real threat. On Dice alone it may tie our model. A free network can use time when it helps, but nothing stops it from predicting smaller infarcts at longer delays in parts of input space it rarely saw. With roughly 150 training cases, such regions are large. We settle this with the counterfactual test below. We sweep $\tau$ for each held-out patient and count monotonicity violations and volume-versus-delay slopes. If the time-channel network is both as accurate and essentially never violates monotonicity, we will report that the structure adds nothing beyond interpretability.
Why not just fit time-dependent perfusion thresholds? This is the traditional clinical extension and costs nothing to deploy. It is included as a baseline. Even so, the challenge gap between thresholding and learned segmentation (0.163 versus 0.285 Dice) suggests that spatial context matters beyond any per-voxel cutoff. If a time-varying threshold closes most of that gap, the learned part of our model is not earning its cost.
Why not use a neural ODE or recurrent trajectory model, as in chronic progression work? ISLES'24 has one follow-up per patient. A learned vector field would be constrained only at two points, and its extra flexibility would mostly fit noise. The closed-form hazard is the two-observation special case with the right monotonicity built in. We would revisit an ODE only with serial perfusion imaging, which this file does not provide.
Proposed Experiments
All experiments use ISLES'24 with patient-level five-fold cross-validation, plus the hidden test set if available. Primary metrics are the challenge's own: Dice, absolute volume difference, and lesion-wise F1, overall and for infarcts under 5 mL.
The first decisive check is delay stratification. We split held-out patients at the median $\tau$ and compare per-stratum Dice and volume error across all models. A model that truly uses time should help most in the long-delay stratum, where static models systematically under-call penumbral loss.
The second check is counterfactual consistency. For each held-out patient, we evaluate predicted infarct volume over $\tau$ from 1 to 24 hours. We report the fraction of patients whose curve ever decreases (zero by construction for our model) and the cohort-average volume-per-hour slope. We compare that slope with the observed cross-patient relation between delay and final volume after adjusting for baseline core. A model whose implied slope disagrees with the data is miscalibrated in time even if its Dice is good.
The third check is the meaning of the residual map. We test whether total predicted residual-death volume $\sum_v r_v$ predicts 3-month mRS (the modified Rankin disability scale) beyond final infarct volume and NIHSS. We also test whether the raw-perfusion-curve arm sharpens that association. A positive result would justify a prospective study with post-treatment perfusion imaging, where no-reflow can be measured directly. A null result means the residual term is absorbing noise and should not be described as no-reflow.
The fourth check is synthetic recovery of planted $c$, $\lambda$, $r$ maps. If the parameters are not identifiable under the real distribution of delays, the decomposition claims are withdrawn and only the accuracy and monotonicity claims stand.
No reported result was found on Lacuna for any time-conditioned or counterfactual infarct prediction on ISLES'24, so the second and third checks have no external number to beat. Their baselines are the ablations above.
References
- ISLES'24 -- A Real-World Longitudinal Multimodal Stroke Dataset
- ISLES'24: Improving final infarct prediction in ischemic stroke using multimodal imaging and clinical data
- Progress in the Treatment of Acute Ischemic Stroke, Current Challenges and the Establishment of Clinical Decision-Making System
- Xiaoxuming Decoction: A Traditional Herbal Recipe for Stroke With Emerging Therapeutic Mechanisms
- Dl-3-n-butylphthalide Reduces Neurovascular Inflammation and Ischemic Brain Injury in Mice
- Nuciferine reduces vascular leakage and improves cardiac function in acute myocardial infarction by regulating the PI3K/AKT pathway
- Experiences about no-reflow in percutaneous transluminal coronary intervention of acute coronary syndrome
- Forecasting Irreversible Disease via Progression Learning
- Surreal-GAN:Semi-Supervised Representation Learning via GAN for uncovering heterogeneous disease-related imaging patterns
- HOPE: Hybrid-Granularity Ordinal Prototype Learning for Progression Prediction of Mild Cognitive Impairment
- Data-Driven Disease Progression Modelling
- BrainODE: Neural Shape Dynamics for Age- and Disease-aware Brain Trajectories
- Deep learning model for individualized trajectory prediction of clinical outcomes in mild cognitive impairment