Features to extract for image segmentation (question)

Follow the full discussion on Reddit.
I'm doing a project for a class I'm taking at University, about x-ray lung segmentation. My idea is to compare 3 different algorithms regrading their performances: Vanilla Autoencoder, U-Net, and XGBoost. The idea is to make a simple comparison, so no need to do extensive augmentations or parameter tunning if the algorithms are already performing well.