README.md

Datasets

📁 DatasetFactory.py

Factory pattern for creating dataset instances from config. The factory searches for dataset classes in the datasets/ folder by looking up the dataset name in the module's namespace.

Creating a New Dataset

For 2D/3D datasets (inherit from BaseDataset):

Required methods:

1. _get_ordered_images_path() → Returns (train_images, train_labels, val_images, val_labels, test_images, test_labels)
2. get_loader(split) → Returns appropriate loader for 'train'/'val'/'test'

Optional: Override _get_patch_loader() if you need a custom weighted sampler for patch-based training.

For 2D sliced datasets (inherit from BaseDataset2DSliced):

Required methods:

1. _get_ordered_images_path() → Returns paths to 3D volumes
2. _extract_slice_indices(split) → Returns list of (volume_idx, slice_idx) tuples
3. get_loader(split) → Returns DataLoader for 2D slices

Examples:

• ATLAS.py: 3D segmentation dataset with optional train/val split files or random splitting
• BraTS2D.py: 2D sliced version extracting axial slices from 3D BraTS volumes
• QaTaCov.py: 2D chest X-ray dataset (preprocessed)
• QaTaCovPreprocess.py: raw/preprocess dataset loader
• QaTaCovTextEmb.py: 2D chest X-ray dataset with report text embeddings

Important Note

When you create a new dataset class, remember to add it to __init__.py in the datasets/ folder. This ensures the DatasetFactory can find and instantiate your custom dataset class. Simply import your new dataset class in the __init__.py file:

from .YourNewDataset import YourNewDataset