EMPAIA Kidney Glomeruli Segmentation App

A demonstrator application for the EMPAIA Ecosystem for Digital Pathology, as featured in Romberg, D., Strohmenger, K., Jansen, C., Küster, T., Weiss, N., Geißler, C., ... & Homeyer, A. (2021). EMPAIA App Interface: An open and vendor-neutral interface for AI applications in pathology. Computer Methods and Programs in Biomedicine, 106596.

The API specification with which this app interacts can be found on EMPAIA Developer Portal

The app wraps a segmentation model developed and trained by Tuguldur Erdene-Ochir (TU-Berlin) and Yuan Xu (TU-Berlin) for the Kaggle challenge HuBMAP: Hacking the Kidney - Identify glomeruli in human kidney tissue images:

[The] challenge is to detect functional tissue units (FTUs) across different tissue preparation pipelines. An FTU is defined as a “three-dimensional block of cells centered around a capillary, such that each cell in this block is within diffusion distance from any other cell in the same block” (de Bo no, 2013). The goal of this competition is the implementation of a successful and robust glomeruli FTU detector. — “HuBMAP - Hacking the Kidney.” Accessed February 28, 2021. https://kaggle.com/c/hubmap-kidney-segmentation.

How to cite

Evans, Theodore, & Karagül, Emirhan. (2022). EMPAIA Kidney Glomeruli Segmentation App (v1.0.0-stable). Zenodo. https://doi.org/10.5281/zenodo.6208423

BibTeX citation details

@software{evans_theodore_2022_6208423,
  author       = {Evans, Theodore and
                  Karagül, Emirhan},
  title        = {EMPAIA Kidney Glomeruli Segmentation App},
  month        = feb,
  year         = 2022,
  publisher    = {Zenodo},
  version      = {v1.0.0-stable},
  doi          = {10.5281/zenodo.6208423},
  url          = {https://doi.org/10.5281/zenodo.6208423}
}

Or refer to Zenodo for more citation options.

Running the app

Requirements

• Internet access
• Docker Engine
• [email protected] (EATS)
• Up to 10 GB storage space for application docker image (additional storage space is required for the EATS services)
• CUDA-enabled GPU (optional)

Option 1 - Pulling the latest Docker image

• Pull the latest application image: docker pull registry.gitlab.com/empaia/examples/kidney-glomeruli-segmentation-app:latest, using this as the <image tag> when registering the app with EATS

Option 2 - Building the Docker image from source

• Clone the project git clone [email protected]:empaia/examples/kidney-glomeruli-segmentation-app.git
• In the project directory, docker build -t <image tag> .

Launching the app with the EMPAIA App Test Suite

• A lightweight testing environment for the EMPAIA ecosystem can be run with the EMPAIA App Test Suite (EATS).
  • Refer to the developer documentation for a step-by-step guide to starting EATS and running the app
  • Download the WSI new_converted_kidney_bigtiles.tif from NextCloud into to the data directory mounted by the EATS
  • You will need to provide the ead.json and the application <image tag> when registering the app
  • Sample EATS input can be found in ./inputs. Alternatively, the app can be parameterised and run from the Workbench Client.

Details

Model

• hacking_kidney_16934_best_metric.model-384e1332.pth
  • single fold/model kaggle LB: 0.873
  • input patch 1024x1024
  • semi supervised UNet with SCSE using Resnet34 as backbone: unet.py
  • example training parameters for supervised learning on 8 GPUs:

  python -m torch.distributed.launch --nproc_per_node 8 train.py --data-root /data/hubmap-kidney-segmentation --jobs=40 \
      --frozen-batchnorm=false --max-epochs=100 --pretrained=imagenet --batch-size-per-gpu=true --batch-size=20 --image-size=1024 --resize=1024 \
      --optim=adamw --learning-rate=3e-4 --weight-decay=0.0 --data-fold=0 --loss-ce=1 --loss-dice=0 --loss-lovasz=0 --apex-opt-level=O2 --sync-bn=True \
      --arch=unet_scse --backbone=Resnet34 \
      --data-aug-image-compression-p=0.3 \
      --data-aug-gauss-noise-p=0.3 \
      --data-aug-gaussian-blur-p=0.3 \
      --data-aug-rgb-aug-p=0.3 \
      --data-aug-color-jitter-p=0.4 \
      --data-aug-rotate-p=0.4 \
      --data-aug-random-scale-p=0.3 \
      --data-aug-clahe-p=0.2 \
      --data-aug-distort-p=0.7
  

Training data

The model was trained on 13 TIFF files of 500MB - 5GB each, with metadata and annotations in polygon and RLE formats

[...] The HuBMAP data used in this hackathon includes 11 fresh frozen and 9 Formalin Fixed Paraffin Embedded (FFPE) PAS kidney images. Glomeruli FTU annotations exist for all 20 tissue samples; some of these will be shared for training, and others will be used to judge submissions. There are over 600,000 glomeruli in each human kidney (Nyengaard, 1992). Normal glomeruli typically range from 100-350μm in diameter with a roughly spherical shape (Kannan, 2019). — “HuBMAP - Hacking the Kidney.” Accessed February 28, 2021. https://kaggle.com/c/hubmap-kidney-segmentation/data.

Available from Kaggle:

tree -L 1 /data/hubmap-kidney-segmentation
/data/hubmap-kidney-segmentation
├── HuBMAP-20-dataset_information.csv
├── sample_submission.csv
├── test
├── train
└── train.csv