AwebNoob/EIU-Net ? reverse-engineered prompt

Build me a Python project for skin lesion segmentation with a U Net style model.

I want it to take dermoscopy images and predict a segmentation mask for the lesion, with a simple training and validation workflow so I can run experiments on my own dataset. Please set it up in a way that feels complete, with the model, dataset loading, loss functions, metrics, and validation code all working together. The network should use the idea from the paper, with improved feature extraction and better skip connections, but keep the code easy to read and change.

Make sure it can load a dataset from folders, train on GPU if available, and report useful segmentation metrics like Dice and IoU. If anything in the paper or current best practice matters, look up the current docs online if you need to. Also include sensible defaults and command line arguments so I can run it without editing a lot of code.