GraphCut + U-Net Segmentation

This repository contains a segmentation project that combines:

• U-Net (deep learning)
• GraphCut (classical / interactive segmentation)

It includes notebooks, source modules, and a pretrained model checkpoint.

---

Important (Regenerating Outputs)

If you want to regenerate outputs/results, use the notebook(s) in Online Notebooks/ and the parameter files in that same folder.

This is the preferred way to reproduce outputs correctly.
Do not rely on random local changes if you want consistent results.

---

Project Structure

• Online Notebooks/ → online notebooks + parameters for regeneration (use this for output reproduction)
• src/models/ → U-Net model code
• src/training/ → training loop
• src/data/ → dataset loader / preprocessing
• src/graphcut/ → GraphCut functions
• train_unet_model.ipynb → U-Net training notebook
• main.ipynb → main experiment notebook
• enhanced_unet_model.pth → pretrained model checkpoint

---

Keep GraphCut Functions (Do Not Delete)

Please do not delete the GraphCut functions.

Instead, keep them and separate/label them clearly so it is easy to see what is:

• U-Net code
• GraphCut code
• Notebook experiment code

This keeps the project organized and preserves your original work.

---

Recommended Usage

To regenerate outputs (recommended)

1. Open the notebook(s) in Online Notebooks/
2. Use the parameter files in that same folder
3. Run cells in order

• The final file is inference_Unet_plus_Graphcut.ipynb that compares UNet and UNet + Graphcut approaches.
• For regenerating UNet only results you can use. inference_enhanced_unet_model.ipynb for enhanced UNet and train_inference_base_unet_model.ipynb (which includes inference code).
• The code used for brain tumor segmentation on private dataset is available in braintumor_segmentation_private_dataset.ipynb but unfortunately we can not put the link to dataset in the nootebook.

To train U-Net locally

Use train_unet_model.ipynb and the modules inside src/.

---

Notess

This repo is meant to preserve both:

• the U-Net pipeline
• the GraphCut pipeline

Please keep both parts documented and clearly separated.