Bi-temporal Image-driven Acute Stroke Evolution Analysis

01

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 a Stavanger University Hospital (SUH) cohort, salvaged and infarcted penumbra showed consistent feature-space separation (Δ̃_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 (Δ̃_cos = 0.460, p < 0.05). Cross-dataset evaluation on ISLES’24 showed similar trends, supporting the consistency of the observed feature-space patterns across cohorts. These findings suggest that admission CTP contains outcome-associated tissue information beyond conventional core-penumbra delineation.

02

Method

Admission CTP (T₁) is registered to follow-up DWI (T₂), and the two label sets are intersected into six outcome-aware ROI classes. Each class pairs what the tissue looked like on admission with the fate it went on to have.

pen_brain penumbra, salvaged
pen_fi penumbra, infarcted
core_brain core, salvaged
core_fi core, infarcted
clb_brain contralateral, healthy
nhb_fi non-hypoperfused, infarcted

survived progressed to final infarct

03

Results

Four region-pair comparisons on the SUH cohort (n = 109) and on ISLES’24 (n = 149), all at patient level. The separation index Δ̃_cos is the median per-patient gap between within-class and between-class cosine similarity of the CNN embeddings; higher means better separated by final tissue state.

Table I · Region-pair comparisons

Feature	Test 1pen_b · pen_fi	Test 2core_b · core_fi	Test 3nhb_fi · pen_fi	Test 4nhb_fi · clb_b
SUH cohort (n = 109)
FE1 · baseline statistics
mean	✓ large	✕ small	✓ small	✓ large
std	✓ med.	✕ small	✕ small	✓ large
skewness	✓ med.	✕ negl.	✕ negl.	✓ large
min	✓ large	✕ small	✕ small	✓ large
max	✓ small	✕ small	✕ negl.	✓ small
kurtosis	✓ large	✕ negl.	✕ negl.	✓ large
FE2 · GLCM radiomics
Imc1	✓ large	✕ small	✕ negl.	✓ large
Imc2	✓ med.	✕ negl.	✓ small	✓ small
Correlation	✓ large	✕ negl.	✓ small	✕ small
MCC	✓ large	✕ small	✓ small	✓ small
Δ̃_cos · FE3 mJ-Net	✓ 0.146	✓ 0.031	✓ 0.127	✓ 0.460
Δ̃_cos · FE4 nnU-Net	✓ 0.091	✓ 0.038	✓ 0.090	✓ 0.193
ISLES’24 (n = 149)
FE1 · baseline statistics
mean	✓ large	✕ negl.	✓ large	✓ large
std	✓ large	✕ negl.	✓ med.	✓ large
skewness	✓ med.	✕ negl.	✕ negl.	✓ large
min	✓ large	✕ negl.	✓ large	✓ large
max	✓ small	✕ negl.	✓ small	✓ negl.
kurtosis	✓ large	✕ negl.	✓ large	✓ large
FE2 · GLCM radiomics
Imc1	✓ med.	✕ negl.	✕ negl.	✓ large
Imc2	✓ large	✕ negl.	✓ large	✓ small
Correlation	✓ med.	✕ negl.	✓ large	✓ small
MCC	✓ small	✕ negl.	✓ large	✕ negl.
Δ̃_cos · FE3 mJ-Net	✓ 0.177	✓ 0.044	✓ 0.126	✓ 0.401
Δ̃_cos · FE4 nnU-Net	✓ 0.140	✓ 0.042	✓ 0.094	✓ 0.115

FE1 and FE2: two-sided Mann–Whitney U with Benjamini–Hochberg FDR (α = 0.05) and Cliff’s d magnitude. FE3 and FE4: median Δ̃_cos with a one-sample Wilcoxon signed-rank test. ✓ significant, ✕ not significant.

Penumbra is where admission imaging carries the most information about eventual fate: salvaged and infarcted penumbra separate across all four feature families, with medium-to-large Cliff’s d for the baseline and GLCM features. Core tissue does not, consistent with advanced ischemic injury already present at baseline. The largest separation throughout is Test 4, suggesting that tissue outside the admission hypoperfused area but later included in the final infarct already differs from healthy tissue at admission. The same pattern holds on ISLES’24 despite differences in centers, acquisition protocols, follow-up interval, and automated rather than manual admission annotations.

t-SNE projections of FE1-FE4 across all six bi-temporal ROI classes. — **Feature spaces.** 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. Class organization is clearer in the CNN embedding spaces than in the handcrafted feature spaces.

t-SNE projections of mJ-Net embeddings for each of the four region-pair tests. — **Region-pair tests.** t-SNE projections of FE3 (mJ-Net) embeddings for each region-pair test on the SUH cohort. Colors denote the two ROI classes being compared.

Bubble plots of median baseline and GLCM feature values per bi-temporal ROI class. — **Phenotypic gradient.** Median baseline (Std, Max, Skewness, Kurtosis) and GLCM (Imc1, Imc2, MCC, Correlation) feature values per bi-temporal ROI class on the SUH cohort, for LVO (filled, larger) and non-LVO (open, smaller) subgroups; bubble size encodes voxel count. A consistent gradient runs from healthy contralateral brain through salvaged penumbra to infarcted tissue.

Table II · Normalisation and pooling ablation

Normalisation · pooling	Test 1pen_b · pen_fi	Test 2core_b · core_fi	Test 3nhb_fi · pen_fi	Test 4nhb_fi · clb_b
mJ-Net
none Mean	0.068	0.018	0.209	0.181
none Median	0.091	0.023	0.213	0.215
none Max	0.033	0.015	0.085	0.119
z-score Mean	0.098	0.029	0.196	0.264
z-score Median	0.114	0.036	0.200	0.290
z-score Max	0.146	0.031	0.127	0.460
nnU-Net
none Mean	0.019	0.009	0.061	0.085
none Median	0.023	0.011	0.074	0.099
none Max	0.011	0.004	0.024	0.051
z-score Mean	0.066	0.029	0.140	0.270
z-score Median	0.071	0.028	0.137	0.259
z-score Max	0.091	0.038	0.090	0.193

Median Δ̃_cos on the SUH cohort; every entry is significant at p < 0.05. Bold marks the best value per test within each model. Z-score normalisation with max pooling was adopted as the single configuration for the reported CNN results.

Table III · Subgroup analysis, mJ-Net

Subgroup	Test 1pen_b · pen_fi	Test 2core_b · core_fi	Test 3nhb_fi · pen_fi	Test 4nhb_fi · clb_b
Stroke hemisphere
Left hemisphere n = 53	✓ 0.133	✓ 0.028	✓ 0.124	✓ 0.466
Right hemisphere n = 51	✓ 0.155	✓ 0.029	✓ 0.143	✓ 0.458
Onset-to-CT time (median split, 1.22 h)
< 1.22h n = 51	✓ 0.163	✓ 0.049	✓ 0.137	✓ 0.453
≥ 1.22h n = 54	✓ 0.137	✓ 0.023	✓ 0.122	✓ 0.448

Separation is consistent across left and right hemisphere strokes and across onset-to-CT timing, so it is not driven by lateralisation or by how early the patient was imaged. Four patients lack onset-to-CT data; bilateral and midline cases are excluded from the hemisphere split.

04

BibTeX

@inproceedings{rahman2026bitemporal,
  title     = {Bi-temporal Image-driven Acute Stroke Evolution Analysis},
  author    = {Rahman, Md Sazidur and Engan, Kjersti and
               Kurz, Kathinka D{\ae}hli and Khanmohammadi, Mahdieh},
  booktitle = {2026 IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI)},
  year         = {2026},
  eprint       = {2602.07535},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV},
  note         = {In press}
}