pathology

Pathology Use Case

Overview

This is a reference app to run infer + train tasks to segment Nuclei. It comes with following 2 pre-trained weights/model (UNET).

• Segmentation Nuclei - This show-cases an example for multi-label segmentation. It tries segment following labels.
  • Neoplastic cells
  • Inflammatory
  • Connective/Soft tissue cells
  • Dead Cells
  • Epithelial
• DeepEdit Nuclei - It is a combination of both Interaction + Auto Segmentation model which is trained to segment Nuclei cells that combines all above labels as Nuclei.
• NuClick - This is NuClick implementation (UNet model) as provided at: https://github.com/mostafajahanifar/nuclick_torch. Training task for monailabel is not yet supported.

Dataset

Above Nuclei models are trained on PanNuke Dataset for Nuclei Instance Segmentation and Classification

Inputs

• WSI Images
• Region (ROI) of WSI Image

Output

Segmentation Mask are produced in one of the following formats

• Standard JSON
• DSA Document (JSON)
• ASAP Annotation XML

Usage (Development Mode)

We recommend to use Digital Slide Archive as endpoint for studies. However you can also also use FileSystem as studies folder.

git clone https://github.com/Project-MONAI/MONAILabel.git
cd MONAILabel
pip install -r requirements.txt

Install Openslide binaries manually and make sure .dll or .so files for openslide are in system load path.

• For windows, make sure <openslide_folder>/bin is added in PATH environment.
• For ubuntu: apt install openslide-tools

FileSystem as Datastore

# download sample wsi image (skip this if you already have some)
mkdir sample_wsi
cd sample_wsi
wget https://demo.kitware.com/histomicstk/api/v1/item/5d5c07539114c049342b66fb/download
cd -

# run server
./monailabel/scripts/monailabel start_server --app sample-apps/pathology --studies datasets/wsi

QuPath

Once the server is up and running, download QuPath and install MONAILabel plugin for QuPath to run annotations using models available through Pathology App.

The current version of plugin comes with limited features to support basic annotations, submitting label and trigger training. Looking for more contributions from QuPath Developer/Opensource community to make it better.

• Make sure MONAILabel Server URL is correctly through Preferences.
• Open Sample Whole Slide Image in QuPath (which is shared as studies for MONAILabel server)
• Add/Select Rectangle ROI to run annotations using MONAI Label models.
• For Interative model (e.g. DeepEdit) you can choose to provide Positive and Negative points through Annotation panel.

Digital Slide Arhive (DSA) as Datastore

DSA

You need to install DSA and upload some test images. Refer: https://github.com/DigitalSlideArchive/digital_slide_archive/tree/master/devops/dsa

MONAILabel Server

Following are some config options:

Name	Description
preload	Preload models into GPU. Default is False.
roi_size	Default ROI Size for inference in [x,y] format. Default is [1024,1024].
dsa_folder	Optional. Comma seperated DSA Folder IDs. Normally it is <folder_id> of a folder under Collections where Images are stored.
dsa_api_key	Optional. API Key helps to query asset store to fetch direct local path for WSI Images.
dsa_asset_store_path	Optional. It is the DSA assetstore path that can be shared with MONAI Label server to directly read WSI Images.

  # run server (Example: DSA API URL is http://0.0.0.0:8080/api/v1)
  ./monailabel/scripts/monailabel start_server --app sample-apps/pathology \
    --studies http://0.0.0.0:8080/api/v1 \

  # run server (Advanced options)
  ./monailabel/scripts/monailabel start_server --app sample-apps/pathology \
    --studies http://0.0.0.0:8080/api/v1 \
    --conf dsa_folder 621e94e2b6881a7a4bef5170 \
    --conf dsa_api_key OJDE9hjuOIS6R8oEqhnVYHUpRpk18NfJABMt36dJ \
    --conf dsa_asset_store_path digital_slide_archive/devops/dsa/assetstore

DSA Client Plugin

You can use projectmonai/monailabel-dsa:latest plugin from dockerhub. Otherwise you can build from sources as follows.

  cd plugins/dsa
  docker build -t projectmonai/monailabel-dsa:latest .

Upload new Task (under Slicer CLI Web Tasks) using the above docker image. This will add/enable MONAILabel under Analysis Page.

Performance Benchmarking

The performance benchmarking is done using MONAILabel server and DSA client. All the details are captured here.

Following is summary of the same:

• NucleiDetection (CPU Based DSA Algorithm)
• Segmentation/DeepEdit (MONAILabel models)