Automated Detection and Quantification of Hemorrhagic Transformation After Endovascular Thrombectomy

This multicenter study demonstrates that an AI-based algorithm accurately detects and quantifies hemorrhagic transformation across multiple imaging modalities following endovascular thrombectomy, showing high sensitivity and a strong dose-response correlation with 3-month functional outcomes that surpasses traditional categorical grading.

The Gist

The Big Picture: A Digital Safety Net for Stroke Patients

Imagine a patient has a major stroke caused by a clogged artery. Doctors use a high-tech "plumbing" procedure called Endovascular Thrombectomy (EVT) to fish out the clot and restore blood flow. It's a life-saving move, but it comes with a tricky side effect: sometimes, when blood rushes back into the damaged brain tissue, it leaks out, causing a hemorrhagic transformation (HT). Think of it like turning on a fire hose to a garden hose that's already cracked; the pressure might cause a new leak.

Doctors need to know immediately: Is there a new leak? How big is it? Is it dangerous?

Traditionally, a human expert has to look at brain scans (like X-rays or MRIs) and manually measure these leaks. But with thousands of patients, this is slow and hard to do perfectly every time.

This paper asks: Can a smart computer (Artificial Intelligence) look at these scans, find the leaks, measure them, and tell us how sick the patient will be, just as well as a human?

---

The Experiment: A Nationwide "Test Drive"

The researchers gathered data from 18 different hospitals across Korea involving 1,490 patients. They tested a "digital detective" (the AI) on three different types of brain scans:

1. NCCT: A standard, quick brain X-ray (like a basic photo).
2. GRE & SWI: Specialized MRI sequences that are super-sensitive to blood (like a high-definition, thermal camera that sees heat leaks).

The AI was trained to spot blood and measure its volume in milliliters. The researchers then compared the AI's findings against the "Gold Standard": a team of expert human doctors who graded the bleeding using a strict rulebook called ECASS.

The Results: The AI is a Super-Sniffer

Here is what the study found, broken down simply:

1. The AI is Great at Finding the "Big Leaks"

The most dangerous type of bleeding is called Parenchymal Hemorrhage (PH). It's a large, space-occupying hematoma that can crush brain tissue.

• The Analogy: If the brain is a house, PH is a room filling up with water.
• The Result: The AI was incredibly good at finding these big leaks. It caught 94% to 98% of them, regardless of whether they used the basic X-ray or the fancy MRI.
• Why it matters: Missing a big leak is dangerous. The AI rarely misses the ones that matter most.

2. The AI is a "Volume Meter," Not Just a "Yes/No" Switch

Old methods just said, "Yes, there is bleeding" or "No, there isn't." The AI, however, gives a specific number: "There is 5 milliliters of blood."

• The Analogy: Imagine a weather forecast. Old methods say, "It might rain." The AI says, "It will rain 2 inches."
• The Result: The researchers found a direct link between the amount of blood the AI measured and how the patient did three months later.
  • Tiny amount (0 mL): 62% of patients had a good recovery.
  • Huge amount (>50 mL): Only 7% had a good recovery; many passed away.
• The Takeaway: The AI's "volume meter" predicts the future better than just saying "bleeding present" or "bleeding absent."

3. The AI Sees What Humans Miss (The "Ghost Leaks")

This was the most surprising part. Sometimes, the expert human doctors looked at the scan and said, "No bleeding here." But the AI said, "Wait, I see a tiny bit of blood (volume > 0)."

• The Result: Even when the humans said "No bleeding," the patients where the AI found a tiny bit of blood did worse than the patients where the AI found absolutely nothing.
• The Analogy: Imagine two houses. One has a visible crack in the wall (Human sees it). The other looks perfect, but the AI detects a tiny, invisible hairline fracture. The house with the invisible fracture actually has a higher risk of collapsing later.
• Why it matters: The AI might be detecting "sub-threshold" damage that human eyes are too tired or the rules are too strict to catch. It suggests the AI is a more sensitive early-warning system.

The "False Alarms" and Limitations

The AI isn't perfect.

• The Confusion: Sometimes, the contrast dye used during the surgery looks like blood on the X-ray. The AI sometimes mistakes this dye for a leak (a "false alarm").
• The MRI Advantage: The AI was even better at spotting leaks on the specialized MRI scans (GRE/SWI) than on the standard X-rays. This is because MRI is like a night-vision camera for blood, while X-rays are like a flashlight in a foggy room.

The Bottom Line: Why This Matters

This study proves that AI can act as a tireless, super-accurate assistant for stroke doctors.

1. Speed: It can scan thousands of images instantly.
2. Precision: It measures the exact size of the bleed, which helps predict how the patient will recover.
3. Safety Net: It might catch tiny, dangerous leaks that human experts miss, allowing for earlier intervention.

In short: The AI isn't replacing the doctor, but it's giving them a powerful new pair of glasses that can see the invisible, measure the unmeasurable, and predict the future with surprising accuracy. This could help doctors make better decisions for stroke patients in the future.

Technical Summary

1. Problem Statement

Hemorrhagic Transformation (HT) is a critical complication following Endovascular Thrombectomy (EVT) for acute ischemic stroke, significantly influencing patient outcomes. Currently, HT is classified manually using categorical systems like the European Cooperative Acute Stroke Study (ECASS) or Heidelberg Bleeding Classification.

• Limitations of Current Methods: Manual grading is labor-intensive, subjective, and impractical for large-scale multicenter registries. Furthermore, existing Artificial Intelligence (AI) algorithms for hemorrhage detection are primarily validated for spontaneous intracranial hemorrhage (ICH) in emergency settings. They have not been rigorously tested in the post-EVT context, where imaging is confounded by iodinated contrast staining, edema, and blood-brain barrier disruption.
• Gap in MRI Analysis: While MRI sequences (GRE and SWI) are more sensitive to blood products than CT, AI application to these sequences for post-EVT HT detection remains unexplored.

2. Methodology

The study was a multicenter diagnostic accuracy analysis conducted within the Clinical Research Collaboration for Stroke in Korea (CRCS-K) registry, involving 18 academic centers.

• Study Population:

  • Cohort: 1,490 patients who underwent EVT for anterior circulation large-vessel occlusion (June 2022–Dec 2023).
  • Inclusion: Follow-up imaging (NCCT, GRE, or SWI) performed within 168 hours (7 days) post-procedure.
  • Modality Distribution: 320 NCCT, 420 GRE, 750 SWI.
  • Reference Standard: Expert adjudication by two vascular neurologists using ECASS criteria (0 to 4). Binary endpoints were defined as Any HT (ECASS ≥1) and Parenchymal Hemorrhage (PH) (ECASS ≥3).

• AI Algorithms:

  • NCCT: Used a commercially available, regulatory-cleared deep learning model (JLK Inc.) trained on spontaneous ICH. It was applied without retraining or domain adaptation.
  • GRE & SWI: Custom segmentation models were developed using the nnU-Net framework (3D fully convolutional encoder-decoder).
    • Training: Trained on multi-vendor datasets (1,441 GRE studies; 1,152 SWI studies) with voxel-level annotations.
    • Post-processing: A separate 2D CSF segmentation model was used to exclude superficial siderosis (overlap ≥50% with CSF) to isolate true parenchymal hemorrhage.
  • Output: The AI provided a binary detection result and a continuous volumetric measurement (mL).

• Statistical Analysis:

  • Performance metrics (Sensitivity, Specificity, AUC) were calculated against the ECASS reference.
  • Optimal volumetric thresholds for PH detection were determined using the Youden Index.
  • Prognostic value was assessed via correlation between AI-derived volume and 3-month modified Rankin Scale (mRS) scores.

3. Key Contributions

1. First Systematic Validation in Post-EVT: This is the first study to validate AI-based hemorrhage quantification specifically for post-EVT HT across three distinct imaging modalities (NCCT, GRE, SWI).
2. Generalizability of Spontaneous ICH Models: Demonstrated that an AI model trained solely on spontaneous ICH (NCCT) retains high diagnostic fidelity in the complex post-EVT setting without domain-specific retraining.
3. MRI-Specific AI Development: Developed and validated dedicated deep learning models for GRE and SWI sequences, achieving high Dice Similarity Coefficients (0.938 for GRE, 0.862 for SWI).
4. Continuous vs. Categorical Prognosis: Proposed that AI-derived continuous volume is a superior prognostic biomarker compared to categorical ECASS grading, revealing dose-response relationships and identifying high-risk subgroups within "negative" expert classifications.

4. Results

• Diagnostic Performance (Parenchymal Hemorrhage Detection):

  • Sensitivity: Exceeded 94% across all modalities.
    • NCCT: 95.4%
    • GRE: 94.4%
    • SWI: 98.3%
  • AUC (Area Under Curve):
    • NCCT: 0.900
    • GRE: 0.943
    • SWI: 0.953
  • Optimal Volume Thresholds for PH: 1.8 mL (NCCT), 1.6 mL (GRE), and 1.6 mL (SWI).

• Prognostic Correlation:

  • AI-derived volume showed a strong positive correlation with 3-month mRS (Spearman ρ = 0.353, P < 0.001).
  • Dose-Response Effect: Good functional outcomes (mRS 0–2) declined sharply as volume increased:
    • 0 mL: 61.8%

    • 50 mL: 6.7%

  • Mortality rates increased correspondingly from 6.2% (0 mL) to 46.7% (>50 mL).

• Discovery of "AI-Positive" Subgroups:

  • Among patients classified as ECASS 0 (No HT) by experts, the AI detected hemorrhage (volume > 0 mL) in 12.8% (112/872) of cases.
  • These "AI-positive/Expert-negative" patients had significantly worse outcomes than true negatives:
    • Good outcome: 48.2% vs. 67.2% (P < 0.001).
    • Mortality: 10.7% vs. 4.6% (P < 0.001).
  • This suggests AI detects sub-threshold pathological changes missed by human categorical grading.

5. Significance and Implications

• Clinical Utility: The high sensitivity for PH (>94%) suggests AI can serve as a reliable triage tool for identifying patients with clinically significant hemorrhage, potentially reducing the burden on radiologists in large registries.
• Prognostic Refinement: The study challenges the sufficiency of categorical grading (ECASS). It demonstrates that AI-derived continuous volume captures prognostic information that categorical systems discard, particularly in the "gray zone" of low-volume hemorrhage.
• Modality Selection: The superior AUC of MRI (GRE/SWI) over NCCT reinforces the use of MRI for follow-up in specialized centers, though NCCT remains highly effective for PH detection.
• Future Directions: The findings support the use of continuous volumetric endpoints in future clinical trials for stroke interventions, offering greater statistical power than ordinal scales. The study also highlights the need for prospective validation of these thresholds in diverse populations.

Conclusion: AI-based hemorrhage quantification provides high-fidelity detection of post-EVT parenchymal hemorrhage across multiple modalities. Crucially, it acts as a continuous prognostic biomarker, identifying at-risk patients even when expert grading suggests no hemorrhage, thereby offering a scalable solution for precision stroke care and research.