Dootmaan/MT-UNet — reverse-engineered prompt

Build me the official MT UNet medical image segmentation project in Python so I can train and test the Mixed Transformer UNet model on ACDC or Synapse datasets.

I want it set up so I can point the training scripts at my local dataset folder, run training on a GPU, save checkpoints, and write prediction outputs to a predictions folder. Include separate runnable paths for ACDC and Synapse like the original project describes. Please make the dataset paths easy to change and include clear command examples for starting training, resuming or loading pretrained weights, and knowing where outputs are saved.

Keep it focused on reproducing the paper code, not a web app. Make sure the model, dataset loading, utilities, and training scripts are organized cleanly. Add a short README that explains how to download the datasets and pretrained weights from the current Zenodo links, how to place the files, and how to run the training. Look up current docs online if you need to.