🤖 AI

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

This paper proposes a bi-temporal analysis framework that uses radiomic and deep learning features from admission CTP and follow-up DWI to characterize stroke evolution, demonstrating that feature embeddings can effectively distinguish between salvageable penumbra and non-salvageable tissue.

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

Published 2026-02-10

📖 3 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a gardener looking at a patch of wilting plants in a large garden. You see some plants that are turning yellow (the penumbra) and some that are already brown and crispy (the ischemic core).

You want to know: “If I water these plants right now, which ones will bounce back to being healthy, and which ones are already too far gone?”

This research paper is essentially trying to build a high-tech "plant diagnostic tool" for the human brain during a stroke.

The Problem: The "Snapshot" Trap

When someone has a stroke, doctors take a quick scan (called CTP) to see which parts of the brain are losing blood. This scan shows two main areas:

The Core: The "dead zone" where tissue is already dying.
The Penumbra: The "danger zone"—tissue that is struggling but might still be saved if doctors act fast.

The problem is that a single scan is just a snapshot in time. It’s like looking at a photo of a person running; you can see they are moving, but you don't know if they are running toward a finish line or toward a cliff. Doctors need to know the trajectory of the tissue.

The Solution: The "Time-Travel" Analysis

The researchers decided to stop looking at just one snapshot. Instead, they looked at two:

Time 1 (T1): The "Emergency Room" scan (the snapshot of the struggle).
Time 2 (T2): The "Follow-up" scan (the final result, showing what actually survived).

By comparing these two, they created six different "neighborhoods" of brain tissue. For example, they looked at "Penumbra that survived" versus "Penumbra that died."

The Method: Looking Beyond the Surface

To understand these neighborhoods, they didn't just look at how bright or dark the scan was. They used three different "magnifying glasses":

The Basic Stats (The "Thermometer"): Looking at simple numbers like average brightness.
The Texture Analysis (The "Fingerprint"): Looking at the patterns and "feel" of the tissue. Is it smooth? Is it grainy? (Like telling the difference between silk and sandpaper).
The AI Deep Dive (The "X-Ray Vision"): They used advanced Artificial Intelligence (Deep Learning) to find hidden patterns that even the human eye can't see. The AI looks at the "vibe" of the data to find subtle clues about whether a cell is truly healthy or just "pretending" to be healthy.

The Big Discovery: The "Hidden Vulnerability"

The most exciting part of their findings is what they discovered about the "Quietly Dying" tissue.

They found a group of tissue that looked perfectly normal on the first scan (the "Not Hypoperfused" group), but later turned into dead tissue. The AI was able to "sense" that this tissue was actually vulnerable, even though it looked fine to the human experts. It’s like seeing a person who looks healthy but has a very high, hidden fever—the AI caught the "fever" before the symptoms showed up.

Why does this matter?

In the real world, this could lead to much smarter stroke treatments. Instead of just treating everyone with a "one size fits all" approach, doctors might one day use these AI tools to say: "This specific patch of brain looks like it has a high chance of recovery, let's prioritize aggressive treatment here," or "This area is already doomed, let's focus our energy elsewhere."

In short: They are teaching computers to read the "future" of brain tissue by studying the subtle patterns of its "present" struggle.

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

1. Problem Statement

Current clinical practice in acute ischemic stroke relies on single-timepoint segmentation of Computed Tomography Perfusion (CTP) to identify the ischemic core (irreversible damage) and the penumbra (at-risk, salvageable tissue). However, these static classifications fail to capture the biological heterogeneity of stroke and the continuous temporal evolution of tissue. There is a clinical need to understand why certain "at-risk" regions successfully reperfuse (salvage) while others progress to infarction. The authors argue that existing methods ignore the latent "tissue phenotypes" that could predict these different evolutionary trajectories.

2. Methodology

The authors propose a bi-temporal analysis framework that characterizes tissue at admission ( $T_1$ ) by linking it to the definitive infarct outcome observed on follow-up Diffusion-Weighted MRI ( $T_2$ ).

Data & ROI Construction: The study uses CTP ( $T_1$ $T_{1}$ ) and DWI ( $T_2$ $T_{2}$ ) from 18 recanalized patients. By intersecting manual segmentations from both timepoints, the authors constructed six distinct Regions of Interest (ROIs) representing different evolutionary paths:
- Salvaged: Healthy tissue ($CLB$), Penumbra $\to$ Brain ( $p_b$ ), and Core $\to$ Brain ( $c_b$ ).
- Infarcted: Non-hypoperfused $\to$ Infarct ( $NHB_{fi}$ ), Penumbra $\to$ Infarct ( $p_{fi}$ ), and Core $\to$ Infarct ( $c_{fi}$ ).
Feature Extraction (Four Modalities):
1. Baseline (BL): First-order statistical descriptors (mean, std, skewness, etc.) using a $3 \times 3 \times 30$ sliding window on 4D CTP data.
2. Radiomics (GLCM): 3D Gray Level Co-occurrence Matrix features (texture metrics like IMC1, IMC2, MCC, and Correlation) to capture spatial heterogeneity.
3. Deep Embeddings (mJ-Net): Spatio-temporal features from a 2D+time CNN encoder designed for stroke segmentation.
4. Deep Embeddings (nnU-Net): Latent features from a standard, high-performance 2D medical segmentation backbone.
Assessment Techniques: The framework uses t-SNE for manifold visualization, Mann-Whitney U tests for statistical significance, Cliff’s delta for effect size, and Cosine Similarity (specifically a "separation index" $\Delta_{cos}$ ) to quantify how well deep features distinguish between salvaged and infarcted tissue.

3. Key Contributions

Bi-temporal Framework: Moving beyond static segmentation to a dynamic "tissue evolution" mapping.
Multi-modal Feature Characterization: Demonstrating that combining statistical, textural, and deep learning-derived features provides a more holistic view of tissue state than any single method.
Latent Vulnerability Discovery: Identifying that regions initially labeled as "not hypo-perfused" ($NHB$) can exhibit feature signatures similar to the ischemic core, suggesting they are biologically vulnerable despite appearing normal on standard perfusion maps.

4. Results

Feature Separability: Both GLCM and Baseline features showed a consistent trend: salvaged tissue (penumbra/healthy) clusters closer to healthy brain phenotypes, while infarcted tissue shows higher dispersion and distinct textural degradation.
Deep Learning Superiority: The mJ-Net architecture provided the strongest separation between salvageable and non-salvageable tissue. The separation index ( $\Delta_{cos}$ ) for the penumbra was statistically significant ( $p = 1.5 \times 10^{-4}$ ), proving that the CNN's latent space effectively encodes the "fate" of the tissue.
Penumbra vs. Core: The analysis found that penumbra regions show significant feature differences depending on their final state (salvaged vs. infarcted), whereas core regions show less significant differentiation, suggesting the penumbra is the most critical zone for phenotyping.
nnU-Net Observations: While nnU-Net showed a clear gradient (Core $\to$ Penumbra $\to$ Healthy), it was less effective than mJ-Net at specifically discriminating between the final tissue states.

5. Significance

This work shifts the paradigm of stroke imaging from segmentation (where is the lesion?) to phenotyping (what is the biological fate of this tissue?). By demonstrating that deep learning encoders capture "tissue phenotypes" that correlate with clinical outcomes, the study provides a foundation for more sophisticated prognostic models. If validated on larger datasets, this approach could help clinicians better identify which patients are most likely to benefit from aggressive reperfusion therapies by looking "beyond the core" into the latent characteristics of the penumbra.