Clinical Research Collaboration for Stroke in Korea Imaging Repository:A Prospective Multicenter Neuroimaging Repository

The CRCS-K Imaging Repository establishes a nationwide, prospective multicenter platform in Korea that integrates comprehensive stroke neuroimaging with AI-based quantification and clinical data, demonstrating the feasibility of large-scale automated analysis to investigate how imaging workflows impact treatment efficiency and patient outcomes.

The Gist

Imagine you are trying to understand how a city handles traffic jams. You could just count the number of cars that get stuck (the "clinical data"), or you could look at the actual video footage of the roads, the weather, the time of day, and the specific behavior of every driver (the "neuroimaging data").

For a long time, stroke research has been like the first option: counting cars. Doctors and researchers have kept excellent records of what happened to stroke patients (did they get a clot-busting drug? did they survive?), but they often threw away the "video footage"—the actual brain scans. They reduced complex, colorful, 3D brain images down to simple checkboxes like "stroke present: yes/no."

This paper introduces a massive new project called the CRCS-K Imaging Repository, which is like building a giant, secure, digital library that stores every single frame of brain scan footage from stroke patients across Korea, along with their medical records.

Here is a breakdown of what they did and why it matters, using some everyday analogies:

1. The "Giant Digital Library" (The Repository)

Think of the hospital system in Korea as a network of 18 different libraries. In the past, if you wanted to study how doctors read brain scans, you'd have to visit each library, ask for a summary, and hope the librarian remembered the details correctly.

The CRCS-K Imaging Repository is a super-library that connects all 18 branches.

• What they collected: They didn't just take a summary. They took the entire movie reel. They collected over 225,000 individual brain scan sequences (CTs, MRIs, angiograms) from nearly 21,000 patients.
• The "Raw" Data: They kept the images in their original, high-definition format (DICOM), just like a filmmaker keeps the raw footage before editing. This means future scientists can re-watch the footage with new tools later, even if those tools didn't exist when the patient was scanned.

2. The "AI Translator" (AISCAN Platform)

Raw video footage is great, but it's hard to search. You can't type "find all brains with a small red spot" into a video player.

Enter AISCAN, the research platform described in the paper. Think of this as a super-smart AI Translator or a Librarian Robot.

• How it works: As soon as a brain scan enters the library, the AI robot watches it. It doesn't just "see" the image; it measures it. It counts the size of the damaged area, measures blood flow, and counts tiny bleeds.
• The Result: It turns the complex, messy pictures into clean, simple numbers (like "lesion size: 5mm"). Now, researchers can search the database using these numbers, just like searching for a book by its ISBN number.

3. The "Traffic Light" Experiment (The Proof-of-Concept)

To show off how useful this library is, the researchers ran a specific experiment. They asked: "Does taking more time to get a detailed brain scan help or hurt the patient?"

In the real world, doctors have a choice:

• Option A (The Express Lane): Do a quick CT scan and treat immediately.
• Option B (The Scenic Route): Do a CT, then a more detailed MRI, then maybe a perfusion scan, and then treat.

What they found:

• The Time Cost: The more "scenic route" steps (more scans) a patient took, the longer they waited for treatment. It's like stopping at every traffic light on the way to the hospital; you get a better view of the city, but you arrive later.
• The Outcome: For patients who needed a mechanical clot removal (EVT), those who took the "Express Lane" (CT-only) seemed to do slightly better than those who took the "Scenic Route" (MRI-heavy).
• The Nuance: The authors are careful to say this doesn't mean MRIs are "bad." It just means that in a real-world emergency, every extra minute spent scanning is a minute not spent saving brain tissue. The "best" path depends on how sick the patient is and how far they are from the hospital.

4. Why This Changes Everything

Before this project, if a researcher wanted to know how different hospitals handled stroke imaging, they had to guess or ask doctors what they usually did.

Now, with this CRCS-K Imaging Repository:

• We can see the truth: We can see exactly how different hospitals behave. Some hospitals are "CT-first" (fast and furious), while others are "MRI-first" (detailed and cautious).
• We can learn from the past: Because the raw images are saved, if a new AI tool is invented in 10 years, scientists can run it on these old images to see if it would have helped those patients.
• It's a living ecosystem: This isn't a static report; it's a growing garden. As new patients arrive, new data is added, and new AI tools can be tested against the whole history of the database.

The Bottom Line

This paper describes the construction of a massive, AI-powered time machine for stroke research. It preserves the raw, detailed reality of how strokes are treated in the real world, rather than just the simplified summaries. By doing this, it allows doctors to ask complex questions like, "Does a 10-minute delay for a better scan actually save more brains, or does it cost more?"—questions that were previously impossible to answer with the data we had.

Technical Summary

1. Problem Statement

Current prospective stroke registries, while valuable for understanding real-world clinical outcomes, suffer from a critical limitation: they typically reduce complex, multidimensional neuroimaging data (CT, MRI, Angiography) into simplified categorical variables (e.g., "site of occlusion" or "presence of hemorrhage"). This simplification forfeits the rich quantitative information inherent in raw imaging studies. Furthermore, there is a lack of large-scale, prospectively collected imaging datasets that are tightly coupled with clinical data and processed through standardized Artificial Intelligence (AI) pipelines. This gap hinders the ability to perform granular analyses on imaging workflows, treatment efficiency, and the precise impact of imaging modalities on patient outcomes.

2. Methodology

The authors established the CRCS-K Imaging Repository, a prospective, multicenter platform built upon the existing Clinical Research Collaboration for Stroke in Korea (CRCS-K) registry.

• Data Collection:

  • Population: 20,792 consecutive acute ischemic stroke patients enrolled from 18 comprehensive stroke centers across South Korea between June 2022 and May 2025.
  • Imaging Scope: The repository captures all neuroimaging performed during the index hospitalization, including Computed Tomography (CT), Magnetic Resonance (MR), and Digital Subtraction Angiography (DSA).
  • Format: Data is collected in native DICOM format with preserved headers. Personally Identifiable Information (PII) is removed, and studies are assigned pseudonymized IDs linked to the clinical registry.
  • Workflow: Participating hospitals transmit data to a central imaging laboratory for quality verification, sequence classification, and metadata extraction.

• AI-Based Quantification:

  • Validated AI models (previously tested in multicenter studies) are applied automatically to eligible images.
  • Outputs: The AI pipeline converts images into standardized numeric features, including ischemic lesion volumes (NCCT/DWI), perfusion parameters, white matter hyperintensity burden, and cerebral microbleed counts.
  • Scalability: The system is designed to allow future re-analysis of the entire archive with new algorithms without re-collecting data.

• Research Platform (AISCAN):

  • An integrated web-based platform (AISCAN) links clinical variables (demographics, risk factors, NIHSS, mRS outcomes) with imaging metadata and AI-derived quantitative features.
  • Access is granted via a formal proposal process, ensuring data security and compliance with Korean regulations.

• Proof-of-Concept Analysis:

  • The authors analyzed the association between pre-treatment imaging workflows and treatment efficiency/outcomes for Intravenous Thrombolysis (IVT) and Endovascular Treatment (EVT).
  • Statistical Approach: Used hierarchical mixed-effects models for time metrics (adjusting for center-level clustering) and propensity score overlap-weighted analyses to assess functional outcomes (modified Rankin Scale), minimizing selection bias.

3. Key Contributions

• Infrastructure: Creation of the first prospective, nationwide repository in Korea that systematically collects all stroke neuroimaging (not just acute triage scans) and integrates it with clinical data.
• AI Integration: Implementation of an automated, standardized quantification pipeline that transforms raw DICOM data into reproducible numeric features at scale, reducing inter-rater variability.
• Granularity: Moving beyond categorical labels to sequence-level data, allowing researchers to analyze the specific composition of imaging protocols (e.g., NCCT vs. CTP vs. DWI).
• Platform: Development of AISCAN, a scalable, secure environment that facilitates federated-style research while maintaining local data governance.

4. Key Results

• Dataset Scale: The repository contains 225,159 imaging sequences from 20,792 patients.
  • Breakdown includes: 37,453 DWI, 35,242 NCCT, 24,829 TOF-MRA, 20,508 SWI, 16,729 GRE, 15,503 CTA, 4,193 CTP, and 5,809 PWI sequences.
• Inter-Hospital Heterogeneity:
  • Significant variation exists in imaging strategies. While CT is the dominant first modality (83.6% overall), the rate of "MR-first" workflows varies drastically across centers (1.0% to 56.7%).
  • MR-first workflows are more common in patients with longer onset-to-arrival times (>12 hours) and lower stroke severity (lower NIHSS).
• Workflow Efficiency:
  • There is a direct, linear relationship between the number of imaging sequences and treatment delays.
  • IVT: Door-to-needle times increased from ~28 minutes (pre-arrival imaging only) to ~79 minutes (MR with perfusion).
  • EVT: Door-to-puncture times increased from ~52 minutes to ~139 minutes with more complex imaging protocols.
• Clinical Outcomes (Proof-of-Concept):
  • EVT: Propensity score analyses suggested numerically more favorable functional outcomes (mRS 0-1/0-2) for patients triaged with CT-based protocols compared to MR-based protocols.
  • IVT: Differences in outcomes between CT and MR protocols were smaller and less consistent.
  • Note: The authors caution that these findings reflect confounding by indication (MR is often used for lower severity/uncertain cases) rather than proving CT is inherently superior.

5. Significance

• Bridging the Gap: The repository successfully bridges the divide between clinical registries and the multidimensional nature of medical imaging, enabling research questions that were previously unanswerable.
• Real-World Evidence: It provides a "natural laboratory" to study the trade-offs between diagnostic precision (more imaging) and treatment timeliness (time-to-treatment) in a real-world setting, free from the rigid protocols of randomized controlled trials.
• Future-Proofing: By retaining raw DICOM files, the repository allows for the retrospective application of future, more advanced AI algorithms, ensuring the data remains valuable as technology evolves.
• Global Model: The architecture serves as a blueprint for how stroke registries can evolve into dynamic, image-rich ecosystems to support the integration of AI into cerebrovascular care, potentially enabling cross-national benchmarking and observational trial emulation.