Skip to content

README.md

Latest commit

 

History

History
80 lines (57 loc) · 2.31 KB

README.md

File metadata and controls

80 lines (57 loc) · 2.31 KB
Raw

Data notes

This public repository does not include the full study dataset.

Included in the repository

  • seg_samples_500/: 500 representative pre-segmented ultrasound image samples organized by fibrosis grade, for repository inspection and demonstration runs.

Not included in the repository

  • the full multi-center dataset used for the main SFibAI experiments;
  • YOLO-format ROI annotation files (optional; only needed for ROI-based cropping — see scripts/preprocessing/extract_roi.py for generating these from raw images);
  • the full 4-class dataset used for baseline re-training;
  • complete private or institution-restricted metadata.

Expected data locations

Main SFibAI pipeline

The default data root is data/seg_samples_500/.

Minimum required structure (images only, sufficient for demo runs):

seg_samples_500/
├── train/              # Training images organized by fibrosis grade (0.0 to 3.4)
│   ├── 0.0/
│   ├── 0.1/
│   └── ...
└── val/                # Validation images organized by fibrosis grade
    ├── 0.0/
    ├── 0.2/
    └── ...

Full structure with ROI annotations (enables crop-based augmentation):

seg_samples_500/
├── train/
├── val/
├── train_label/        # (optional) YOLO-format .txt annotation files for training images
└── val_label/          # (optional) YOLO-format .txt annotation files for validation images

Each .txt annotation file corresponds to an image file with the same stem name, in the format:

class_id x_center y_center width height

where coordinates are normalized to [0, 1].

When annotation files are absent, the dataset loader automatically disables cropping and loads images directly.

Tip: If you have raw ultrasound images with device-UI backgrounds, you can generate both segmented ROI images and the corresponding YOLO annotation files automatically:

python scripts/preprocessing/extract_roi.py --help

Reproduced baselines

The repository defaults point to:

data/baseline_data/cls_raw4/

Expected structure:

cls_raw4/
├── train/
├── val/
├── train_label/
└── val_label/

The Lee VGG baseline uses only the image folders. The Guo SVM baseline also requires the ROI annotation files.