whisney/DoseDiff — reverse-engineered prompt

Build me a working Python project that reproduces the DoseDiff radiotherapy dose prediction workflow from the paper and makes it easy to run on the OpenKBP dataset.

I want an end to end setup where someone can put the provided OpenKBP data into the repo, run the preprocessing to turn the raw files into the training format, generate the distance maps and missing ROI data, train the diffusion model, and then run prediction and evaluation on saved model weights. Please keep the existing scripts but wire everything up so the steps are clear, paths are consistent, and it actually runs without guesswork. If anything is missing or outdated, look up current docs online and fix it.

Please also make the README simple and practical, with the exact order of commands, what each step produces, and any assumptions about GPUs, checkpoints, or file locations. If there are common failure points, handle them or document them clearly.