Brain Tumor Detector — U-Net vs U-MambaNet (MONAI/PyTorch)

A reproducible 3D MRI brain-tumor segmentation pipeline comparing a strong U-Net baseline with a U-MambaNet (state-space) variant.
Includes training, inference, and a gallery of overlays (ground truth vs. predictions).

---

🔎 Overview

• Goal: segment brain tumors on 3D MRI (BraTS-style) and compare U-Net vs U-MambaNet fairly.
• Stack: PyTorch + MONAI, AMP (mixed precision), sliding-window training/inference.
• Report: Bearcat AI Grant mid-project update (Nov 2025).

Repo map

Repo map (key files)
├─ train.py               # main training entry
├─ engine.py              # training/eval loop utils
├─ model_builder.py       # U-Net wrapper
├─ U_Mamba_net.py         # U-MambaNet model
├─ U_Mamba_blocks.py      # SSM/Mamba blocks
├─ U_Net_predict.py       # inference: U-Net
├─ U_Mamba_predict.py     # inference: U-MambaNet
├─ preprocess.py          # orientation/spacing + intensity norm
├─ RepeatChannel.py       # channel utilities
├─ requirements.txt
└─ assets/
   ├─ figs/               # overlays & screenshots
   └─ media/              # demo .gif/.mp4 (tracked with Git LFS)

---

🚀 Quick Start

# 1) create env (example)
conda create -n tumor python=3.10 -y && conda activate tumor
pip install -r requirements.txt

# 2) train (choose your model)
python train.py --model unet         # or --model u_mamba
# add your flags: --epochs, --roi, --batch, --amp, etc.

# 3) run inference
python U_Net_predict.py       --input <path_to_volume> --output outputs/
python U_Mamba_predict.py     --input <path_to_volume> --output outputs/

Ground Truth (Case 112)

U-Net Pred (Case 112)

U-MambaNet Pred (Case 112)

🎬 Demo

Dang Luu, "Brain Tumor Detector — U-Net vs U-MambaNet," 2025.

---

Tips for great visuals

• Consistency: same slice index and color map across GT/U-Net/U-Mamba panels.
• Dimensions: keep PNGs ≤ 1400px width and GIFs ≤ 15s to avoid heavy pages.
• Alt text: set good alt text in <img> for accessibility.
• Relative paths: use assets/... so images work on all branches/forks.
• ITK-SNAP screenshots: hide toolbars/cursors; export with a neutral background; consider overlay opacity 0.4–0.6 and a consistent label palette.