Beyond Core and Penumbra: Bi-Temporal Image-Driven Stroke Evolution Analysis

TL;DR

This study proposes a bi-temporal framework to characterize ischemic brain tissue by linking admission-time CTP signatures ($T_1$) to follow-up DWI outcomes ($T_2$) using six bi-temporal ROIs. It integrates statistical, radiomic, and deep CNN features from two encoders ($mJ$-Net and nnU-Net) and evaluates tissue evolution via clustering and embedding similarity, revealing pronounced separation for penumbra that salvages versus infarct and strong salvageability signals in deep embeddings, especially from $mJ$-Net. The findings support the feasibility of imaging-based tissue phenotyping for stroke evolution and highlight the potential for improved prognosis and treatment planning, while acknowledging limitations like small sample size and single-center scope. Future work aims for larger multicenter validation and exploration with public datasets to generalize the bi-temporal phenotyping framework.

Abstract

Computed tomography perfusion (CTP) at admission is routinely used to estimate the ischemic core and penumbra, while follow-up diffusion-weighted MRI (DWI) provides the definitive infarct outcome. However, single time-point segmentations fail to capture the biological heterogeneity and temporal evolution of stroke. We propose 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 admission (T1) and post-treatment follow-up (T2). All features are extracted at T1 from CTP, with follow-up DWI aligned to ensure spatial correspondence. Manually delineated masks at T1 and T2 are intersected to construct six regions of interest (ROIs) encoding both initial tissue state and final outcome. 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. Both baseline GLCM and deep embeddings showed a similar trend: penumbra regions exhibit features that are significantly different depending on final state, whereas this difference is not significant for core regions. Deep feature spaces, particularly mJ-Net, showed strong separation between salvageable and non-salvageable tissue, with a penumbra separation index that differed significantly from zero (Wilcoxon signed-rank test). These findings suggest that encoder-derived feature manifolds reflect underlying tissue phenotypes and state transitions, providing insight into imaging-based quantification of stroke evolution.

Figures (6)

• Figure 1: Framework of the proposed method. Acute ischemic regions from CTP at $T_1$ are combined with final tissue outcome from DWI at $T_2$ to form six ROI classes. Image-derived features from four extraction methods are computed on CTP and grouped according to ROI membership.
• Figure 2: Construction of the tissue bi-temporal masks by combining manual delineation at admission ($T_1$) and follow-up ($T_2$). At $T_1$, preprocessed 4D CTP is annotated into brain, penumbra and core. At $T_2$, DWI is mapped to CTP slab and DWI co-registered to CTP is annotated for final infarct. The $T_1$ and $T_2$ masks are then merged into six-ROI mask.
• Figure 3: Baseline and GLCM feature extraction. (a) Baseline features are computed using a $3\times3\times30$ sliding window on CTP and aggregated into region-wise vectors via tissue evolution masks. (b) Radiomic-based 3D GLCM features are extracted per slice and assigned to each tissue evolution class.
• Figure 4: CNN-based deep feature extraction. Left: the 2D+time mJ-Net encoder processes $16\times16\times30$ patches and applies global average pooling on the last convolutional block to yield a 256D feature vector per patch. Right: the nnU-Net encoder produces a $256\times64\times64$ feature map, providing a 256D feature vector at each position of a downsampled $64\times64$ grid.
• Figure 5: Box plots for GLCM and BL features over the six ROIs on slice level.