Federated Learning (AI-Vengers Assemble)

Phase 1: CXR

This document outlines the data and computational requirements for participating in the AI-Vengers Federated Learning effort.

Compute Requirements

1. Deep Learning Base AMI (Ubuntu 18.04)
2. Test/Demo - t2.2xlarge
3. 8 vCPU, 32GB memory, 100GB EBS disk, network performance “Moderate”
4. At least 1 GPU

Data Requirements

Mandatory Fields:

• Self Reported Race
• Gender
• Age

Optional Fields:

• Scanner Type
• Finding vs No Finding

Label Format

0 - American Indian or Alaska Native
1 - Asian
2 - Black or African American
3 - Hispanic or Latino
4 - Native Hawaiian or Other Pacific Islander
5 - White

Data Preparation

1. Extract metadata and pngs from your DICOMs using Niffler PNG Extaction found here https://github.com/Emory-HITI/Niffler/tree/master/modules/png-extraction
2. Prepare a CSV file with two columns ["Path", "Race"]
3. Run the data preprocessing scripts which will output a JSON file for the train, test, and validation sets
   python create_jsons.py [path/to/csv]
4. Copy these JSON files into the data directory of your client
5. Symlink your JSON file to a file named dataset.json ln -s [json file] dataset.json