melisklc0/stroke-classification

Stroke Classification: Pre-processed Brain CT Dataset

Related resources

• Live demo
• Training code
• Distilled model

Summary

• Task: Binary CT image classification (No-Stroke = 0, Stroke = 1)
• Dataset Size: ~16.5k images per fold, ~49.5k files across folds on disk. ~3 GB total
• Structure: 3-fold cross-validation (Fold1, Fold2, Fold3), each containing train/ and test/ splits
• Labeling Strategy: Original classes (No Stroke, Bleeding, Ischemia) were converted into a binary setup by merging Bleeding and Ischemia into a single Stroke class for emergency triage

Preprocessing Pipeline

1. Merged Bleeding and Ischemia into a unified Stroke class.
2. Applied stratified 3-fold cross-validation with separate train/ and test/ sets.
3. Balanced test sets to ~750 images per class per fold.
4. Applied data augmentation on training sets:
   • Rotation (±10°)
   • Zooming
   • Translation
   • Horizontal flipping resulting in ~7,500 images per class per fold.

Repository Structure

The dataset is organized to support cross-validation training workflows and external generalization testing seamlessly:

• dataset/: Contains all pre-processed, augmented, and validation CT scan images.
  • External_Dataset/: An independent dataset (sourced from Kaggle) used purely for external validation and testing model generalization across different scanner calibrations.
  • Fold1/: Training and validation split 1.
  • Fold2/: Training and validation split 2.
  • Fold3/: Training and validation split 3.

Inside each fold and the external dataset, the images are categorized into two sub-folders: Stroke and No-Stroke.
Train on FoldN/train, validate on FoldN/test. Rotate folds for 3-fold cross-validation.

Source Data & Attribution

The data utilized to build this augmented dataset originates from two primary sources. In compliance with open data policies, please find the attributions below:

1. Primary Dataset (Turkish Ministry of Health): The raw CT scans utilized to build the core augmented training folds were obtained from the Open Data Portal of the Republic of Turkey Ministry of Health.
   • Original Source: Open Data Portal of the Turkish Ministry of Health
2. External Validation Dataset (Kaggle): An external dataset was used purely to test the robust generalization of the trained models across different scanner calibrations.
   • Original Source: Head CT Hemorrhage Dataset by Felipe Kitamura

Licensing & Usage

This processed dataset is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to share and adapt this material for any purpose, even commercially, under the following terms:

• Attribution: You must give appropriate credit to the original data sources (mentioned above) and indicate that this is an augmented, pre-processed version of the original files.

Disclaimer: This dataset is provided for research and educational purposes. It is not intended to replace professional medical diagnosis or serve as a standalone clinical tool.

Authors

• Melis Kılıç
• Esra Koç

Advisor: Assoc. Prof. Dr. Kali Gürkahran