Pathology Search Comparison

This pipeline is comprised of three main steps:

1. Downloading and storing database and test datasets.
2. Running each of the four models' to create model specific database and generating results.
3. Post-Hoc analysis of the results.

To reproduce our results, you can follow the guidelines presented here. Also, please feel free to use our pipeline to test these models on you custom database and test datasets.

Throughout this repository, you will often find folders with .gitignore inside them. These folders are just placeholders to demonstrate how the directory often looks like after you run the model. Due to the large size of these compiled data, it was not possible to include them in this repository. However, they are available upon reasonable request.

Setting Up the Conda Environment:

We recommend having Mamba installed. You can run the following command to install all the dependencies:

mamba/conda env create -f search.yml

Downloading Data

We have provided dtailed guidelines on how to download the TCGA slides we used as a common database, and test slides we used for different studies in the data_preparation subdirectory of this repository. Please refer to data_preparation/README.md for detailed explanation.

Running the Models

For each of the four methods we compared in this study, we have provided the source code we used to generate the results in this repository in the following subdirectories:

1. yottixel
2. sish
3. retccl
4. hshr

Inside each one of these subdirectories, you can find a detailed explanation on how to run these models in their corresponding README.md file.

Post-Hoc Analysis

The final aggregation and analysis of the results is performed using scripts provided here in the analysis subdirectory. A brief explanation of how to run these scritps is provided in analysis/README.md.

Refrences

This study has evaluated the following refrences for readiness to be deployed in the clinical settings:

1. Yottixel
2. SISH
3. RetCCL
4. HSHR