Bi-temporal Image-driven Acute Stroke Evolution Analysis

The gist The proposed framework explores differences in hypoperfused areas that end up being either infarcted or re-perfused and can be seen as tissue characterization

Bi-temporal Analysis Framework

The work proposes a bi-temporal analysis framework that characterizes ischemic tissue using statistical descriptors, radiomic texture features, and deep feature embeddings from two architectures (mJ-Net and nnU-Net) Bi-temporal refers to the admission (T1) and the first post-treatment follow-up (T2) All features are extracted at T1 from CTP, and follow-up DWI is aligned with CTP to ensure spatial correspondence Manually delineated masks at T1 and T2 are intersected to construct six regions of interest encoding both initial tissue state and final outcome Extracted features were aggregated per region and analyzed in feature space Evaluation on 18 patients with successful reperfusion demonstrated meaningful clustering of region-level representations Regions classified as penumbra or healthy at T1 that ultimately recovered exhibited feature similarity to preserved brain tissue, whereas infarct-bound regions formed distinct groupings

Feature Extraction Methods

The study considers three feature families to characterize ischemic tissue These include baseline statistical descriptors, radiomic texture metrics, and deep CNN embeddings which together capture complementary aspects of tissue heterogeneity and provide a compact representation of stroke evolution between the two time points

1. Baseline features (FE1) are derived from 3D CTP data by computing six first order statistics—mean, standard deviation, skewness, kurtosis, minimum and maximum—using a sliding window W(¯x) in the CTP volumes These features are aggregated by element–wise max pooling into a single descriptor F BL pt,ROI,zi∈ R 6

2. Radiomic features (FE2) capture higher order spatial texture beyond first order intensity statistics by extracting 3D gray level co–occurrence matrix (GLCM) features Four descriptors are retained: IMC1 and IMC2, MCC, and correlation

3. Deep CNN-based embeddings (FE3 & FE4) are derived from two segmentation encoders trained on the cohort (excluding the 18 recanalized LVO patients) These include a 2D+time mJ-Net and a 2D nnU-Net framework

Assessment Techniques

The assessment techniques used to evaluate the feature characterization include several methods

[a] Box plots summarize the distribution of baseline statistical and GLCM radiomic features across bi–temporal ROIs, providing a compact view of central tendency, variability and class–level overlap to qualitatively assess feature separability

[b] t-SNE visualization qualitatively assess the geometric structure of feature representations by projecting the high-dimensional feature vectors into two-dimensional space

[c] Region-wise feature discriminability is quantitatively assessed using non-parametric statistical tests, specifically the MannWhitney U test with Bonferroni correction applied to pairwise comparisons

[d] Cosine similarities quantify alignment in feature space using absolute cosine similarity defined as the absolute value of the cosine of the angle between embedding vectors

Key Findings

A consistent trend emerges across both GLCM and BL features, where healthy tissue that remains viable (ROIb CLB) exhibits high median values, narrow interquartile ranges, and low variability In contrast, core tissue evolving into final infarct (ROIf i c) shows lower values and greater dispersion Among GLCM descriptors Imc2 provides the clearest separation between healthy/penumbra and infarcted tissue The mJ-Net exhibits more distinct partitioning between salvaged and infarct-prone tissue among all four Test 1 is always significant, through all types of features, pointing to a clear difference in penumbra being salvaged and penumbra ending as infarct Overall, our findings support the feasibility of using acute CTP-derived feature spaces as a basis for imaging phenotyping of stroke tissue evolution

Limitations

The study is limited by cohort size which restricts statistical power and generalizability Another limitation arises from co-registration between CTP and DWI, as the two modalities differ in acquisition geometry This introduces an axial angulation mismatch that standard 2D registration cannot fully compensate for Ischemic lesions may remain unstable beyond the 24-72 hours follow-up window which could affect the definition of final infarct regions and influence the interpretation of the reported results

Conclusion and Future Work

The proposed framework opens avenues for larger validation and for introducing stroke tissue phenotypes into stroke analysis Future work will focus on extending the cohort size and we will try to evaluate our work on multicenter public datasets The study is therefore positioned as the methodological feasibility analysis, intended to evaluate the proposed feature characterization The proposed framework opens avenues for larger validation and for introducing stroke tissue phenotypes into stroke analysis

Ethical Compliance

This study was conducted in accordance with institutional guidelines and approved by the regional ethical committee in Norway, and all data were pseudonymised prior to analysis This study was conducted in accordance with institutional guidelines and approved by the regional ethical committee in Norway, and all data were pseudonymised prior to analysis<ref:

1. Bi-temporal Tissue Characterization for Stroke Progression: The improved system can characterize ischemic tissue using statistical descriptors, radiomic texture features, and deep feature embeddings from two architectures (mJ-Net and nnU-Net) to map regions into six distinct ROI classes encoding both initial tissue state and final outcome.

2. Predictive Phenotyping via Deep Feature Manifolds: The system can leverage encoder-derived feature manifolds may reflect underlying tissue phenotypes and state transitions to assess salvageability, as shown by the finding that Deep feature spaces, particularly mJ-Net, showed strong separation between salvageable and non-salvageable tissue.

3. Quantification of Tissue Vulnerability: The AI system can quantify subtle changes in tissue vulnerability by analyzing how features describe evolution, specifically noting that None-hypoperfused at admission but infarct-evolving regions has features that are closer to core/infarct suggests that learned representations may capture latent vulnerability not apparent in initial expert labels.

4. Discriminative Texture Analysis for Core vs. Penumbra: The system can utilize GLCM features, specifically Imc2 provides the clearest separation between healthy/penumbra and infarcted tissue, allowing for more robust discrimination than raw intensity values alone.

5. State Transition Classification via CNN Embeddings: By using deep feature embeddings, the system can perform clustering patterns in feature space, enabling a clear observation of state transitions, such as the finding that Clear state transition from core, penumbra to healthy tissue was seen specially in nnU-Net.

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 (Δ =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 (Δ =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.