Bi-temporal Image-driven Acute Stroke Evolution Analysis

Md Sazidur Rahman; Kjersti Engan; Kathinka Dæhli Kurz; Mahdieh Khanmohammadi

cs.CVFeb 7, 2026

Bi-temporal Image-driven Acute Stroke Evolution Analysis

Authors: Md Sazidur Rahman, Kjersti Engan, Kathinka Dæhli Kurz, Mahdieh Khanmohammadi

Organizations: Department of Electrical Engineering and Computer Science, University of Stavanger, Norway · Stavanger Medical Imaging Laboratory, Stavanger University, Hospital, Norway

Abstract

Acute ischemic stroke requires rapid treatment decisions that are strongly guided by emergency imaging. Admission computed tomography perfusion (CTP) is commonly used to estimate the ischemic core, representing irreversibly damaged tissue, and the penumbra, representing hypoperfused but potentially salvageable tissue. Follow-up diffusion-weighted MRI (DWI) is then used to define the final infarct. Existing imaging approaches primarily focus on core--penumbra segmentation or final-infarct prediction, but provide limited insight into how heterogeneous penumbral tissue evolves after treatment. We propose a bi-temporal tissue phenotyping framework that links admission CTP signatures with follow-up DWI-defined tissue outcome using six outcome-aware region-of-interest classes. Admission tissue signatures are characterized using statistical, radiomic, and deep-learning features extracted from mJ-Net and nnU-Net representations. On an internal cohort (SUH), salvaged and infarcted penumbra showed consistent feature-space separation ( $\tildeΔ{\cos}=0.146$ , $p<0.05$ ), while core tissue showed minimal separation by subsequent fate. The largest separation was observed between initially non-hypoperfused tissue that later infarcted and healthy contralateral tissue ( $\tildeΔ{\cos}=0.460$ , $p<0.05$ ). Cross-dataset evaluation on the publicly available ISLES'24 dataset showed similar trends, supporting the consistency of the observed feature-space patterns. These findings suggest that admission CTP contains outcome-associated tissue information beyond conventional core-penumbra delineation. The code is available at https://github.com/yokko123/bi-temporal-ctp-dwi-code.

Figures & tables

Panel of Figure 1 — Fig. 1 : Framework Overview. ( Top ): bi-temporal ROI construction, intersecting registered admission CTP and follow-up DWI masks into six outcome-aware classes. ( Bottom ): CTP feature extraction (FE1–FE4). $\text{ROI}_{x}^{y}$ , admission class $x$ and outcome $y$ : $c$ core, $p$ penumbra, NHB non-hypoperfused, CLB contralateral, $b_{T_{2}}$ non-infarcted, $fi$ final infarct; NCCT non-contrast CT.

Characteristic	SUH (after preproc.)	ISLES’24
Patients (analysis)	109/149	149/149
Age, y (mean $\pm$ SD)	71.9 $\pm$ 14.4	71.1 $\pm$ 14.7
Admission NIHSS	9.4 $\pm$ 8.3	12.0 $\pm$ 6.3
Imaging ( $T_{1}$ $\rightarrow$ $T_{2}$ )	CT $\rightarrow$ MRI	CT $\rightarrow$ MRI
Median Size (NCCT grid)	$512\times 512\times 50$	$512\times 581\times 69$
CTP temporal res.	1 s	1 s

TABLE I : Summary of the SUH and ISLES’24 cohorts.

ROI class	SUH (n=109)	ISLES’24 (n=149)
core_fi	1,451 [212–7,832]	2,450 [347–8,444]
core_brain	1,569 [268–5,933]	977 [286–4,476]
pen_fi	6,924 [996–17,953]	11,594 [4,715–41,232]
pen_brain	70,189 [17,834–155,738]	170,273 [91,339–279,119]
clb_brain	467,646 [251,024–647,513]	934,954 [763,774–1,060,942]
NHB_fi	3,216 [851–8,992]	9,792 [3,452–27,560]

TABLE II : Median voxels per patient [IQR] by ROI class.

	Test 1	Test 2	Test 3	Test 4
Feature	$\text{ROI}_{p}^{b}$ vs $\text{ROI}_{p}^{fi}$	$\text{ROI}_{c}^{b}$ vs $\text{ROI}_{c}^{fi}$	$\text{ROI}_{\textit{NHB}}^{fi}$ vs $\text{ROI}_{p}^{fi}$	$\text{ROI}_{\textit{NHB}}^{fi}$ vs $\text{ROI}_{\textit{CLB}}^{b}$
SUH cohort ( $n=109$ )
FE1
Mean	✓ large	$\times$ small	✓ small	✓ large
Std	✓ med.	$\times$ small	$\times$ small	✓ large
Skewness	✓ med.	$\times$ negl.	$\times$ negl.	✓ large

TABLE III : Region-pair comparisons. FE1/FE2: Mann–Whitney U (FDR,

\alpha{=}0.05

), Cliff’s

d

effect size. FE3/FE4: median

\tilde{\Delta}_{\cos}

, one-sample Wilcoxon. Symbol = significance ( green :✓/

\textbf{red}:\times

), shade = effect-size magnitude (Cliff’s

d

0.147/0.33/0.474

;

\tilde{\Delta}_{\cos}

0.1/0.3

Fig. 2 : t-SNE projections of FE1 - FE4 across all six bi-temporal ROI classes on the SUH cohort. Colors denote tissue fate and markers denote ROI classes.

Fig. 3 : Bubble plots of median BL (Std, Max, Skewness, Kurtosis) and GLCM (Imc1, Imc2, MCC, Correlation) feature values per bi-temporal ROI class on the SUH cohort for LVO (filled circles, larger) and non-LVO (open markers, smaller) patient subgroups. Bubble size encodes voxel count. A consistent phenotypic gradient from healthy ( $\text{ROI}_{\textit{CLB}}^{b}$ ) through salvaged penumbra ( $\text{ROI}_{p}^{b}$ ) to infarcted tissue is visible across both subgroups.

Fig. 4 : t-SNE projections of FE3 (mJ-Net) embeddings for each region-pair test on the SUH cohort: Test 1 ( $\text{ROI}_{p}^{b}$ vs $\text{ROI}_{p}^{fi}$ ), Test 2 ( $\text{ROI}_{c}^{b}$ vs $\text{ROI}_{c}^{fi}$ ), Test 3 ( $\text{ROI}_{\textit{NHB}}^{fi}$ vs $\text{ROI}_{p}^{fi}$ ), Test 4 ( $\text{ROI}_{\textit{NHB}}^{fi}$ vs $\text{ROI}_{\textit{CLB}}^{b}$ ). Colors denote the two ROI classes being compared in each test.

			Test 1	Test 2	Test 3	Test 4
Model	Norm.	Aggr.	$\text{ROI}_{p}^{b}$ vs $\text{ROI}_{p}^{fi}$	$\text{ROI}_{c}^{b}$ vs $\text{ROI}_{c}^{fi}$	$\text{ROI}_{NHB}^{fi}$ vs $\text{ROI}_{p}^{fi}$	$\text{ROI}_{NHB}^{fi}$ vs $\text{ROI}_{CLB}^{b}$
mJ-Net	None	Mean	✓ 0.068	✓ 0.018	✓ 0.209	✓ 0.181
		Median	✓ 0.091	✓ 0.023	✓ 0.213	✓ 0.215
		Max	✓ 0.033	✓ 0.015	✓ 0.085	✓ 0.119
	Z-score	Mean	✓ 0.098	✓ 0.029	✓ 0.196	✓ 0.264
		Median	✓ 0.114	✓ 0.036	✓ 0.200	✓ 0.290

TABLE IV : Ablation: effect of Z-score normalization and pooling aggregation on CNN embedding separation (

\tilde{\Delta}_{\cos}

, one-sample Wilcoxon, SUH cohort). All entries significant (

p<0.05

). Bold = highest per test within model.

		Test 1	Test 2	Test 3	Test 4
Subgroup	$n$	$\text{ROI}_{p}^{b}$ vs $\text{ROI}_{p}^{fi}$	$\text{ROI}_{c}^{b}$ vs $\text{ROI}_{c}^{fi}$	$\text{ROI}_{NHB}^{fi}$ vs $\text{ROI}_{p}^{fi}$	$\text{ROI}_{NHB}^{fi}$ vs $\text{ROI}_{CLB}^{b}$
Stroke hemisphere
Left	53	✓ 0.133	✓ 0.028	✓ 0.124	✓ 0.466
Right	51	✓ 0.155	✓ 0.029	✓ 0.143	✓ 0.458
Onset-to-CT time (median split: 1.22h)
$<1.22$ h	51	✓ 0.163	✓ 0.049	✓ 0.137	✓ 0.453

TABLE V : Subgroup analysis of mJ-Net embedding separation (

\tilde{\Delta}_{\cos}

, one-sample Wilcoxon, SUH cohort). ✓ = significant;

\times

= not significant.

Explore similar work

StrokeTimer: Robust Representation Learning for Ischemic Stroke Onset-Time Estimation from Non-contrast CT
Jun 3, 2026Weiru Wang, Susanne G. H. Olthuis, Elizaveta Lavrova +4Medical Image Analysis Medical Image Representation Learning
Department of Biomedical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands · Graduate School of Life Sciences, Utrecht University, Utrecht, The Netherlands · Department of Neurology, Maastricht University Medical Centre+, Maastricht, The Netherlands +3
EPRA U-Net: An Efficient Pyramid Residual Attention Framework for Accurate Infarct Segmentation in Diffusion-Weighted MRI
Jul 3, 2026Hasan Ulutas, Muhammet Emin Sahin, Mustafa Fatih Erkoc +4U-Net Magnetic Resonance Imaging
Yozgat Bozok University, Department of Computer Engineering, Yozgat, Turkey · Izmir Bakircay University, Department of Computer Engineering, Izmir, Turkey · Queen Mary’s Digital Environment Research Institute (DERI), London, UK +3
Ischemic Stroke Segmentation and Net Water Uptake Quantification on Multicenter Non-Contrast CT Using Supervised Target-Domain Adaptation
Sep 17, 2026Linus Britt, Maximilian Nielsen, Susan Klapproth +5Medical Image Segmentation
Institute for Applied Medical Informatics, University Medical Center Hamburg-Eppendorf, Hamburg, Germany · Institute of Computational Neuroscience, University Medical Center Hamburg-Eppendorf, Hamburg, Germany · Department of Diagnostic and Interventional Neuroradiology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany +4

Jun 3, 2026, cs.CVJ/K move · Enter open · S save

StrokeTimer: Robust Representation Learning for Ischemic Stroke Onset-Time Estimation from Non-contrast CT

Weiru Wang, Susanne G. H. Olthuis, Elizaveta Lavrova, Robert J. van Oostenbrugge+3

Department of Biomedical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands · Graduate School of Life Sciences, Utrecht University, Utrecht, The Netherlands · Department of Neurology, Maastricht University Medical Centre+, Maastricht, The Netherlands+3

Medical Image Analysis Medical Image Representation Learning Computed Tomography

Open paper AskPDF arXiv

Ischemic stroke is a major global disease. Treatment decisions are highly time-sensitive, as eligibility for reperfusion therapies relies on the interval between stroke onset and intervention. However, the true onset time is often uncertain in clinical practice, necessitating imaging-based assessment of tissue age as a surrogate marker. Early ischemic changes on routinely acquired non-contrast CT (NCCT) are often subtle, and real-world clinical datasets exhibit pronounced onset-time class imbalance and center-scanner-related heterogeneity. In this work, we propose StrokeTimer, a fully automated framework for onset-time estimation in acute ischemic stroke. StrokeTimer integrates self-supervised disentanglement learning with energy-guided contrastive learning to capture subtle ischemic patterns while addressing long-tailed data distributions under acquisition variability. Onset time is categorized into three clinically relevant windows: <4.5 h, 4.5-6 h, and >6 h. Experimental results on a large multi-center NCCT dataset from two national cohorts, MR CLEAN Registry and MR CLEAN LATE, show that StrokeTimer achieves a macro AUC of 0.69 and a macro F1-score of 0.57, improving the strongest baseline by nearly 50% (p < 0.005). In this realistic, challenging setting, representative baseline approaches exhibit near-chance macro performance. Model explanations further highlight subtle gray-white matter blurring and hypodense regions consistent with established radiological biomarkers. These findings demonstrate the potential of StrokeTimer to support treatment decision-making in acute ischemic stroke. Code is available at https://github.com/BrainVas/StrokeTimer.

Abstract

Explore similar work

StrokeTimer: Robust Representation Learning for Ischemic Stroke Onset-Time Estimation from Non-contrast CT

EPRA U-Net: An Efficient Pyramid Residual Attention Framework for Accurate Infarct Segmentation in Diffusion-Weighted MRI

Ischemic Stroke Segmentation and Net Water Uptake Quantification on Multicenter Non-Contrast CT Using Supervised Target-Domain Adaptation