Tiger Submission For Leaderboard 2

This repository contains the algorithm used to submit to tigerchallenge leaderboard 2 under the team name SRI.
The algorithm can be found here.
The algorithm explaination can be found here.
The model weights for the algorithm can be found here shared under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

Requirements

• Ubuntu software

  • Ubuntu20.04
  • ASAP 2.0

• Python packages

  • numpy==1.21.5
  • tqdm==4.62.3
  • torch==1.10.2
  • torchvision==0.11.3
  • scikit-image==0.19.2
  • opencv-python==4.5.5.64
  • albumentations==1.1.0

Summary of the files in package

The packages consist of the following python files.

__init__

This is an empty file used for the initialization of the package directory.

__main__

Contains code for calling the package as a module. Runs the process function from the processing file.

gcio

Contains code that deals with grand challenge input and output. It includes predefined input and output paths.

rw

Contains code for reading and writing. Includes function for reading a multi resolution image. Furthermore, it includes classes for writing required files for the challenge, namely: segmentation mask file, detection JSON file, and TILS score file.

processing

Main processing file. Includes code for processing a slide and applies process functions to generate TILS score. The segmentation and detection JSON files generated are not valid and generated only for the purpose of submission.

saved_models

bihead_cell_tissue\resnet34 contains the saved model for the TILS regressor which is a bihead network with Resnet34 as its backbone. tumorbed contains the saved model for tumorbed binary classifier with Resnet18 as its backbone.

utils

This directory contains helper functions for post processing and test time augmentations and TILS regressor model

TIL_score

This file contains the main code where TIL score is calculated based on the patches collected while processing a whole slide image. It is assumed all the patches at a uniform stride can be stored in RAM for a single slide. For some slides there may be out of memory issues. Please contact the author/ raise an issue if that happens

trainingcodes

Please read the README.md specified in the file

Setup

A simple and minimal setup file is included to install the package via pip. Note that the package is not in the PyPI repository.

Dockerfile

Dockerfile to be build and uploaded to grand-challenge. It installs

• nvidia/cuda:11.1-devel-ubuntu20.04
• python3.8-venv,
• ASAP2.0,
• tigeralgorithmexample + requirements

As an entry point, the __main__ file will be run; hence process function from the processing file will be called.

Test and Export

To test if your algorithm works and (still) produces the correct outputs you add an image to ./testinput/ and a corresponding tissue mask in ./testinput/images/

After the image and the tissue background are present in the test and test/images folder, you can run the following command to build and test the docker:

./test.sh

This will build the docker, run the docker and check if the required output is present. Furthermore, it will check if the detected_lymphocytes.json is in valid json format. When there are no complaints in the output you can export the algorithm to an .tar.xz file with the following command:

./export.sh

The resulting .tar.xz file can be uploaded to the grand-challenge platform