Dataset Generation¶

GlassCut provides two approaches for working with tile datasets:

DatasetGenerator¶

The DatasetGenerator processes multiple slides and persists tiles, thumbnails, and metadata to disk with checkpoint/resume support.

from glasscut import DatasetGenerator, GridTiler

tiler = GridTiler(
    tile_size=(512, 512),
    magnification=20
)

generator = DatasetGenerator(
    dataset_id="my_dataset",
    output_dir="./output",
    tiler=tiler,
    n_workers=4,
    batch_size=128,
)

slide_paths = ["slide1.svs", "slide2.svs", "slide3.svs"]
metadata = generator.process_dataset(slide_paths)

print(f"Processed {metadata.total_slides} slides, {metadata.total_tiles} tiles")

Output Structure¶

output/
└── my_dataset/
    ├── slide_001/
    │   ├── tiles/          # Extracted tile PNGs
    │   ├── thumbnails/     # Slide thumbnails
    │   ├── masks/          # Tissue masks
    │   └── slide_metadata.json
    ├── slide_002/
    │   └── ...
    ├── metadata.json       # Dataset-level metadata
    └── processed.json      # Checkpoint file (for resume)

LiveSlideDataset¶

The LiveSlideDataset keeps tiles in memory for interactive exploration and prototyping.

from glasscut import LiveSlideDataset, GridTiler

tiler = GridTiler(
    tile_size=(512, 512),
    magnification=20
)

dataset = LiveSlideDataset(
    slide_paths=["slide1.svs", "slide2.svs"],
    tiler=tiler,
)

sample = dataset[0]  # LiveSlideSample
print(f"Slide: {sample.slide_name}")
print(f"Tiles: {len(sample.tiles)}")