wsinfer 0.3.2

Run patch-based classification on pathology whole slide images.

WSINfer: deep learning inference on whole slide images

Original H&E	Heatmap of Tumor Probability

🔥 🚀 Blazingly fast pipeline to run patch-based classification models on whole slide images.

Table of contents

• Available models
• Installation
  • Pip
  • Containers
    • Containers for different classification tasks
  • Developers
• Examples
  • Setup directories and data
  • On "bare metal" (not inside a container)
  • Run in an Apptainer container (formerly Singularity)
  • Run in a Docker container
  • Output
  • Convert to GeoJSON (for QuPath and other viewers)
  • Convert to Stony Brook QuIP format
  • Add your own model

Available models

Classification task	Output classes	Model	Weights name	Reference
Breast adenocarcinoma detection	no-tumor, tumor	inceptionv4	TCGA-BRCA-v1	ref
Breast adenocarcinoma detection	no-tumor, tumor	resnet34	TCGA-BRCA-v1	ref
Breast adenocarcinoma detection	no-tumor, tumor	vgg16mod	TCGA-BRCA-v1	ref
Lung adenocarcinoma detection	lepidic, benign, acinar, micropapillary, mucinous, solid	resnet34	TCGA-LUAD-v1	ref
Pancreatic adenocarcinoma detection	tumor-positive	preactresnet34	TCGA-PAAD-v1	ref
Prostate adenocarcinoma detection	grade3, grade4+5, benign	resnet34	TCGA-PRAD-v1	ref
Tumor-infiltrating lymphocyte detection	til-negative, til-positive	inceptionv4nobn	TCGA-TILs-v1	ref

Installation

Pip

Pip install this package from GitHub. First install torch and torchvision (please see the PyTorch documentation). We do not install these dependencies automatically because their installation can vary based on a user's system. Then use the command below to install this package.

python -m pip install --find-links https://girder.github.io/large_image_wheels wsinfer

To use the bleed edge, use

python -m pip install \
    --find-links https://girder.github.io/large_image_wheels \
    git+https://github.com/kaczmarj/patch-classification-pipeline.git

Containers

Use the Docker / Singularity / Apptainer image, which includes all of the dependencies and scripts. See DockerHub for the available tags.

• Apptainer / Singularity

  Replace apptainer with singularity if you do not have apptainer

  apptainer pull docker://kaczmarj/patch-classification-pipeline
  

• Docker

  docker pull kaczmarj/patch-classification-pipeline
  

Containers for different classification tasks

We distribute containers that include weights for different tasks, and these containers have a simplified command-line interface of command SLIDE_DIR OUTPUT_DIR. See DockerHub for the available tags. The Dockerfiles are in dockerfiles/ Here is an example:

apptainer pull docker://kaczmarj/patch-classification-pipeline:v0.2.0-paad-resnet34
CUDA_VISIBLE_DEVICES=0 apptainer run --nv --bind $(pwd) patch-classification-pipeline_v0.2.0-paad-resnet34.sif \
    --wsi-dir slides/ --results-dir results/

Developers

Clone this GitHub repository and install the package (in editable mode with the dev extras).

git clone https://github.com/kaczmarj/patch-classification-pipeline.git
cd patch-classification-pipeline
python -m pip install --editable .[dev] --find-links https://girder.github.io/large_image_wheels

Examples

Here we demonstrate running this pipeline on a sample image. Before going through this, please install the package (see Installation).

Setup directories and data

We make a new directory to store this example, including data and results. Enter the following commands into a terminal. This will download a sample whole slide image (170 MB). For this example, we only use one whole slide image, but you can apply this pipeline to an arbitrary number of whole slide images — simply put them all in the same directory.

mkdir -p example-wsi-inference
cd example-wsi-inference
mkdir -p sample-images
cd sample-images
wget -nc https://openslide.cs.cmu.edu/download/openslide-testdata/Aperio/CMU-1.svs
cd ..

List available models and weights

We use "model" as in architecture (like "resnet50"), and "weights" are the pretrained parameters that are loaded into the model for a particular task (like "TCGA-BRCA-v1" for breast cancer tumor detection). Use the following command to list all available models and weights.

wsinfer list

On "bare metal" (not inside a container)

Run the pipeline (without a container). This will apply the pipeline to all of the images in sample-images/ (only 1 in this example) and will write results to results/. We set CUDA_VISIBLE_DEVICES=0 to use the first GPU listed in nvidia-smi. If you do not have a GPU, model inference can take about 20 minutes.

CUDA_VISIBLE_DEVICES=0 wsinfer run \
    --wsi-dir sample-images/ \
    --results-dir results/ \
    --model resnet34 \
    --weights TCGA-BRCA-v1 \
    --num-workers 8

Run in an Apptainer container (formerly Singularity)

I use the commands apptainer here, but if you don't have apptainer, you can replace that with singularity. The command line interfaces are the same (as of August 26, 2022).

apptainer pull docker://kaczmarj/patch-classification-pipeline

Run the pipeline in Apptainer.

CUDA_VISIBLE_DEVICES=0 apptainer run \
    --nv \
    --bind $(pwd) \
    --pwd $(pwd) \
    patch-classification-pipeline_latest.sif run \
        --wsi-dir sample-images/ \
        --results-dir results/ \
        --model resnet34 \
        --weights TCGA-BRCA-v1 \
        --num-workers 8

Run in a Docker container

First, pull the Docker image.

docker pull kaczmarj/patch-classification-pipeline

This requires Docker >=19.03 and the program nvidia-container-runtime-hook. Please see the Docker documentation for more information. If you do not have a GPU installed, you can use CPU by removing --gpus all from the command below.

We use --user $(id -u):$(id -g) to run the container as a non-root user (as ourself). This way, the output files are owned by us. Without specifying this option, the output files would be owned by the root user.

When mounting data, keep in mind that the workdir in the Docker container is /work (one can override this with --workdir). Relative paths must be relative to the workdir.

Note: using --num_workers > 0 will require a --shm-size > 256mb. If the shm size is too low, a "bus error" will be thrown.

docker run --rm -it \
    --shm-size 512m \
    --gpus all \
    --env CUDA_VISIBLE_DEVICES=0 \
    --user $(id -u):$(id -g) \
    --mount type=bind,source=$(pwd),target=/work/ \
    kaczmarj/patch-classification-pipeline run \
        --wsi-dir sample-images/ \
        --results-dir results/ \
        --model resnet34 \
        --weights TCGA-BRCA-v1 \
        --num-workers 2

Output

This will create the following directory structure

results/
├── masks
├── model-outputs
├── patches
└── stitches

• masks contains PNGs of tissue masks
• model-outputs contains CSVs of model outputs
• patches contains HDF5 files of patch coordinates
• stitches contains PNGs with patches stitched together

The output also contains a file results/run_metadata.json containing metadata about the run.

Convert to GeoJSON (for QuPath and other viewers)

GeoJSON is a standardized format to represent geometry. The results of model inference are a type of geometric data structure. Popular whole slide image viewers like QuPath are able to load labels in GeoJSON format.

wsirun togeojson results/ geojson-results

Convert to Stony Brook QuIP format

The Stony Brook QuIP format uses a combination of JSON and plain text files. Provide a unique --execution-id that identifies this run. An example could be tcga-brca-resnet34-tumor. Also provide a --study-id, which could be TCGA-BRCA. The option --make-color-text will create the color-* files that contain color information for each patch in the input slides. This option is disabled by default because it adds significant processing time.

wsirun tosbu \
    --wsi-dir slides/ \
    --execution-id UNIQUE_ID_HERE \
    --study-id TCGA-BRCA \
    --make-color-text \
    --num-processes 16 \
    results/ \
    results/model-outputs-sbubmi/

Add your own model

Define a new model with a YAML configuration file. Please see the example below for an overview of the specification.

# The version of the spec. At this time, only "1.0" is valid. (str)
version: "1.0"
# Models are referenced by the pair of (architecture, weights), so this pair must be unique.
# The name of the architecture. We use timm to supply hundreds or network architectures,
# so the name can be one of those models. If the architecture is not provided in timm,
# then one can add an architecture themselves, but the code will have to be modified. (str)
architecture: resnet34
# A unique name for the weights for this architecture. (str)
name: TCGA-BRCA-v1
# Where to get the model weights. Either a URL or path to a file.
# If using a URL, set the url_file_name (the name of the file when it is downloaded).
# url: https://stonybrookmedicine.box.com/shared/static/dv5bxk6d15uhmcegs9lz6q70yrmwx96p.pt
# url_file_name: resnet34-brca-20190613-01eaf604.pt
# If not using a url, then 'file' must be supplied. Use an absolute or relative path. If
# using a relative path, the path is relative to the location of the yaml file.
file: path-to-weights.pt
# Size of patches from the slides. (int)
patch_size_pixels: 350
# The microns per pixel of the patches. (float)
spacing_um_px: 0.25
# Number of output classes from the model. (int)
num_classes: 2
# Names of the model outputs. The order matters. class_names[0] is the name of the first
# class of the model output.
class_names:  # (list of strings)
  - notumor
  - tumor
transform:
  # Size of images immediately prior to inputting to the model. (int)
  resize_size: 224
  # Mean and standard deviation for RGB values. (list of three floats)
  mean: [0.7238, 0.5716, 0.6779]
  std: [0.1120, 0.1459, 0.1089]

Once you save the configuration file, you can use it with wsinfer run:

wsinfer run --wsi-dir path/to/slides --results-dir path/to/results --config config.yaml