DRAW models are trained using nn-UNet - a UNet based automatic segmentation system that automatically adapts to the dataset provided and selects an appropriate segmentation pipeline for the dataset. This is an open source package which has been released under an Apache 2.0 licence. See the GitHub page https://github.com/MIC-DKFZ/nnUNet for more details on nn-UNet.
Datasets used for training are first manually curated by experienced radiation oncologists. These imaging datasets are typically taken from patients who have provided consent for image banking under the CHAVI project (https://chavi.ai). Datasets are reviewed and structure names are harmonised to meet the TG-263 standards. Automatic segmentation templates are created which include multiple models. Each model can have one or more structures / regions of interest which will be trained for automatic segmentation. We generally have a hold out sample of 10 - 15% of the test cases for model performance evaluation.
Model training is performed on Paperspace machines so that multiple models can be trained together. After training, the model performance is validated on the hold out samples and quantitative model metrics calculated. Experienced oncologists will also review the model segmentation manually. After this is completed, the models are made available to the general public. Model datasets will also be made available on https://chavi.ai.