ULS23 challenge container

This repository contains code that allows one to run nnUNet models trained on images of size 64x128x128 on the ULS23 challenge on Grand Challenge. It takes the challenge's images, crops them to 64x128x128, runs inference with given model weights, then pads the results back to 128x256x256 for submission to the challenge. This framework was used for ULS+, the next iteration of the ULS model. More information about ULS+, along with model weights, can be found here: https://github.com/DIAGNijmegen/oncology-uls-plus.

• /architecture/extensions/nnunetv2 contains the extensions to the nnUNetv2 framework that should be merged with your local install.
• /architecture/input/ contains an example of a stacked VOI image and the accompanying spacings file. Uncommenting line 63 and 66 in the Dockerfile will allow you to run your algorithm locally with this data and check whether it runs inference correctly.
• /process.py is where the model is loaded, predictions are made and postprocessing is applied. If you're testing this model locally and want to use a CPU instead of a GPU, you can do this by changing 'cuda' to 'cpu' in line 23 of process.py.