Modular online patch streaming from whole-slide images for computational pathology.

Stream patches directly from WSIs during training — no disk pre-extraction, no storage overhead. Every component is pluggable: backends, tissue detectors, samplers, filters, transforms, views, dataset adapters.

Install

pip install "wsistream[openslide]"   # with OpenSlide
pip install "wsistream[tiffslide]"   # with TiffSlide (pure Python)
pip install "wsistream[torch]"       # add PyTorch integration (WsiStreamDataset, DDP)
pip install "wsistream[all]"         # everything (OpenSlide + TiffSlide + PyTorch + albumentations + matplotlib)

For development:

git clone https://github.com/RamonKaspar/wsistream.git
cd wsistream
pip install -e ".[dev]"

Documentation

Full documentation: ramonkaspar.github.io/wsistream

To build locally:

pip install mkdocs-material
mkdocs serve          # local preview at http://127.0.0.1:8000

How it works

Each slide goes through a fixed pipeline:

1. Open slide: via an explicit backend (OpenSlideBackend or TiffSlideBackend)
2. Detect tissue: run a TissueDetector on a low-res thumbnail to get a binary mask
3. Sample coordinates: a PatchSampler proposes (x, y) locations within tissue regions
4. Extract patch: read the pixel data from the slide at each coordinate
5. Filter patch: a PatchFilter accepts or rejects the tile based on its pixels
6. Transform patch: apply augmentations (HEDColorAugmentation, RandomFlipRotate, etc.) or produce named multi-view outputs
7. Yield result: PatchResult with image (single-view) or named views (multi-view), coordinates, tissue fraction, and metadata

Quick start

from wsistream.pipeline import PatchPipeline
from wsistream.backends import OpenSlideBackend
from wsistream.tissue import CLAMTissueDetector
from wsistream.sampling import RandomSampler
from wsistream.filters import HSVPatchFilter
from wsistream.transforms import ComposeTransforms, HEDColorAugmentation, RandomFlipRotate, ResizeTransform
from wsistream.datasets import TCGAAdapter

pipeline = PatchPipeline(
    slide_paths="/data/tcga",  # directory or list of files
    backend=OpenSlideBackend(),
    tissue_detector=CLAMTissueDetector(),
    sampler=RandomSampler(patch_size=256, num_patches=-1, target_mpp=0.5),
    patch_filter=HSVPatchFilter(min_pixel_fraction=0.6),
    transforms=ComposeTransforms(transforms=[
        HEDColorAugmentation(sigma=0.05),
        RandomFlipRotate(),
        ResizeTransform(target_size=224),
    ]),
    dataset_adapter=TCGAAdapter(),
    pool_size=8,
    patches_per_slide=100,
    cycle=True,
)

for result in pipeline:
    print(result.image.shape)                # (224, 224, 3) uint8
    print(result.coordinate.mpp)             # ~0.5
    print(result.tissue_fraction)            # 0.87
    print(result.slide_metadata.patient_id)  # TCGA-3L-AA1B

Pool-based slide interleaving

The pipeline keeps pool_size slides open simultaneously and takes patches_per_slide patches from each before closing it and opening the next. With cycle=True, slides are re-queued for infinite streaming. Set patches_per_visit (default 1) to read multiple patches from the same slide before round-robining, which can significantly improve I/O throughput on network filesystems.

PyTorch integration

wsistream.torch provides WsiStreamDataset (an IterableDataset), MonitoredLoader for throughput tracking, and partition_slides_by_rank for DDP. Worker-level slide partitioning is handled automatically.

from torch.utils.data import DataLoader
from wsistream.backends import OpenSlideBackend
from wsistream.sampling import RandomSampler
from wsistream.tissue import OtsuTissueDetector
from wsistream.torch import WsiStreamDataset, partition_slides_by_rank

my_slides = partition_slides_by_rank("/data/tcga", rank=rank, world_size=world_size)

dataset = WsiStreamDataset(
    slide_paths=my_slides,
    backend=OpenSlideBackend(),
    tissue_detector=OtsuTissueDetector(),
    sampler=RandomSampler(patch_size=256, num_patches=-1, target_mpp=0.5),
)

loader = DataLoader(dataset, batch_size=64, num_workers=4, pin_memory=True)
loader_iter = iter(loader)

for step in range(total_steps):
    batch = next(loader_iter)
    images = batch["image"].to(device, non_blocking=True)  # (B, 3, H, W) float32