数据集获取与加载（Hugging Face datasets）

# 前置依赖: pip install datasets
from datasets import load_dataset

ds = load_dataset("Karrar-Alhdrawi/brain-stroke-ct-dataset")
print(ds)  # 查看 splits 与字段结构
# 国内网络可先设镜像: import os; os.environ["HF_ENDPOINT"]="https://hf-mirror.com"

git lfs install && git clone https://hf-mirror.com/datasets/Karrar-Alhdrawi/brain-stroke-ct-dataset

数据集大小约 1.60 GB。

仓库文件结构

• .gitattributes
• README.md
• app.py
• brain-stroke-dataset-v1.zip

数据集卡片摘录（源站说明）

AI-Based Intelligent Extraction of Diseased Areas from Medical Radiography Images

1. Overview

This dataset is a refined and preprocessed version of the Brain Stroke CT Dataset, specifically optimized for semantic segmentation tasks using deep learning architectures like U-Net. This work is part of a research project focused on automating the extraction of pathological regions (Bleeding and Ischemia) from head CT scans to assist in rapid clinical diagnosis.

2. Source & Attribution

The raw data was originally sourced from the Brain Stroke CT Dataset on Kaggle, which was prepared by radiologists and supported by the Turkish Health Institutes (TUSEB).

3. Preprocessing & Methodology

The original dataset provided raw CT images and “Overlay” images where specialists manually highlighted diseased areas using specific colors (Red and Green). To make this data compatible with high-precision segmentation models, a custom Python pipeline was developed to:

• Binary Mask Extraction: Used HSV color masking to isolate the clinicians’ annotations from the overlays, converting them into pure binary masks (White: Diseased, Black: Healthy).
• Normal Case Synthesis: Generated corresponding empty (all-black) masks for healthy brain images to maintain consistency during model training.
• Structural Standardization: Organized the data into a clean images/ and masks/ hierarchy for each category (Normal, Bleeding, Ischemia).

4. Why This Custom Pipeline?

The raw Kaggle dataset is excellent for classification, but lacks ready-to-use binary masks for pixel-wise segmentation. By applying this custom script:

1. We standardized the labels for U-Net ingestion.
2. We ensured that the model learns the exact morphology of the stroke rather than just its general location.
3. We reduced data noise by filtering out non-essential metadata and raw DICOM formats, focusing on standardized PNG outputs.

5. Implementation Script

The following script was used to generate this processed version:

# [
import os
import shutil
import cv2
import numpy as np
from tqdm import tqdm
import kagglehub
from huggingface_hub import HfApi, create_repo

# --- إعدادات الحساب والروابط ---
my_token = "***********"
repo_name = "iraqigold/brain-stroke-ct-dataset"

# مسارات الشغل بداخل الكولاب
raw


## 数据加载示例（图像类）

```python
from PIL import Image
import glob, os

files = (glob.glob(os.path.join(path, "**", "*.png"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.jpg"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.tif"), recursive=True))
print("图像文件数:", len(files))
img = Image.open(files[0]); print("尺寸/模式:", img.size, img.mode)

# torchvision Dataset 方式:
# from torchvision import datasets
# ds = datasets.ImageFolder(path)  # 要求 子目录=类别

目录组织与标注格式以源站说明和下载后实际文件为准。

完整数据卡：huggingface.co/datasets/Karrar-Alhdrawi/brain-stroke-ct-dataset