Stroke Triage With CT and MRI Models

ISEF Category: Biomedical and Health Sciences

Subcategory: Pathophysiology  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

A stroke clock starts fast, but not every patient with weakness or confusion has a true stroke. Some cases are stroke mimics, and that first decision can change treatment. You can build a model that sorts the two using public scan data and NIHSS-style text features. Then you can test whether its confidence is honest, not just high.

What Is It?

A stroke mimic is a condition that looks like stroke at first but turns out to be something else, like a seizure, migraine, low blood sugar, or infection. That matters because doctors need quick triage decisions, and the wrong call can send a patient down the wrong path. Your project asks whether a computer model can help with that first pass.

The model combines two kinds of clues. The images show what the brain scan looks like, while NIHSS-style text features act like a checklist from the bedside exam. Think of it like solving a mystery with both a photo and a short witness statement. The big question is not only whether the model is right, but whether its confidence matches reality.

Why This Is a Good Topic

This makes a strong science fair topic because you can test it with public data, clear labels, and measurable outcomes. It connects to a real clinical problem, since triage speed and error rates matter in stroke care. You can also learn real research skills like data cleaning, patient-level splits, model comparison, and calibration checks without needing a wet lab.

Research Questions

• How does adding NIHSS-style text features change stroke-mimic classification compared with imaging alone?
• What is the effect of probability calibration on the reliability of high-confidence stroke predictions?
• Does a multimodal model perform better than an imaging-only model on public stroke datasets?
• To what extent do different calibration methods reduce overconfident false positives?
• Which feature set best separates stroke mimics from true strokes under class imbalance?
• How does performance change when you train on one dataset and test on another?

Basic Materials

• Laptop with at least 16 GB RAM.
• Python 3.11 installed.
• Jupyter Notebook or VS Code.
• Access to public CT or MRI datasets such as AISD and ISLES.
• Spreadsheet software for tracking experiments and results.
• Reliable internet connection for downloading data and papers.
• Cloud storage or an external drive for large files.

Advanced Materials

• GPU workstation with a CUDA-capable NVIDIA card.
• DICOM-compatible storage and viewing software.
• Secure encrypted drive for protected data handling.
• 3D Slicer or ITK-SNAP for image inspection and annotation.
• Institutional access to clinical stroke image archives.
• Statistical software for calibration plots and decision-curve analysis.

Software & Tools

• Python: Runs your data cleaning, model training, and evaluation scripts.
• Jupyter Notebook: Lets you document experiments and compare results in one place.
• scikit-learn: Handles baseline classifiers, metrics, and calibration checks.
• PyTorch: Supports custom multimodal neural network models.
• MONAI: Provides medical imaging tools for CT and MRI workflows.
• Google Colab: Gives you free or low-cost GPU access for faster training.

Experiment Steps

1. Define the target labels and decide how you will separate true strokes, stroke mimics, and uncertain cases.
2. Choose one imaging-only baseline, then build a second version that also uses NIHSS-style text features.
3. Plan a patient-level split so the same case never appears in both training and testing sets.
4. Build a calibration check so you can compare predicted confidence against actual outcomes.
5. Set up error analysis to see which mimic types the model confuses most.
6. Compare results across datasets or sites to see whether the model still works outside the training data.

Common Pitfalls

• Training and testing on slices from the same patient, which inflates accuracy without real generalization.
• Using raw confidence scores as if they were calibrated probabilities, which hides overconfident wrong calls.
• Mixing MRI and CT records without a clear modality plan, which lets the model learn dataset quirks instead of triage signals.
• Ignoring class imbalance, which can make the model look good by favoring the larger true-stroke class.
• Skipping error review on mimic subtypes, which leaves you blind to patterns such as seizure, migraine, or metabolic cases.

What Makes This Competitive

A class-level version stops at accuracy. A stronger version tests whether the model stays reliable on held-out data, then checks calibration, sensitivity, and false alarm cost separately. You get extra strength if you compare imaging-only and multimodal models, use patient-level splits, and analyze which mimic types trigger the biggest errors. That turns the project into a study of trustworthy triage support, not just a classifier.

Project Variations

• Use CT only instead of CT plus MRI, and test whether the simpler setup still separates mimics from true strokes.
• Swap NIHSS-style text features for basic triage notes, then compare which text signal helps most.
• Focus on calibration by comparing Platt scaling, isotonic regression, and no calibration across the same model.

Learn More

• PubMed: Search review articles on stroke mimics, multimodal medical imaging, and model calibration.
• NIH Stroke Scale resources: Find official training and scoring material from the National Institute of Neurological Disorders and Stroke.
• MONAI documentation: Read the open-source medical imaging toolkit docs for MRI and CT model workflows.
• scikit-learn documentation: Use the calibration and classification sections for probability checks and baseline models.
• 3D Slicer: Explore the open-source imaging platform docs for viewing and preparing DICOM data.