No More Slice Wars: Just code

This repository contains the code and configuration files used for our submission to the Brain Tumor Segmentation Challenge (BraTS 2025), Task 8.

0. System requirements

This code was tested on ubuntu 22.04 LTS and CUDA 12.6, but will work on Windows too.

1. Repository Structure

BraSyn25/
├── final_eval/ # Folder for the final local evaluation
│   └── sbatchable_scripts/     # Sbatch files specific for evaluation
├── slurm/                      # Examples of sbatch files used during training
│   ├── sbatchable_scripts/
│   └── segmenter_script/
├── src/
│   ├── data/                   # Dataset code, collate functions and stuff
│   ├── models/                 # Pytorch models
│   ├── training/               # Training scripts
│   └── utils/                  # Utilities
└── README.md                   # This file! :)

2. Data Preparation

This year's challenge used 3 different datasets:

• Glioma.
• Metastasis.
• Metastasis - add (additional samples).

After downloading, prepare the 3d original dataset following this structure:

work_path/
└── brats3d/
    ├── train_gli/                  
    ├── tran_met/                 
    ├── train_met_add/              
    ├── val_gli/              
    └── val_met/                 

Remember to edit the file paths in the code you are using! In the validation dataset, segmentations are missing. We obtained good quality segmentations, used only during the training phase, using nnunet. You can obtain them using the newly released brats package too. Make sure to get those segmentations before jumping to the next step. After downloading the dataset, you will have to slice it, running src/data/dataset_slicer.py.
This will create the 2d dataset directories, following the sctructure:

work_path/
├── brats2d_axi/
│   ├── train_gli/                   
│   ├── tran_met/                    
│   ├── train_met_add/               
│   ├── val_gli/             
│   └── val_met/     
├── brats2d_cor/
│   ├── train_gli/                   
│   ├── tran_met/                    
│   ├── train_met_add/               
│   ├── val_gli/             
│   └── val_met/  
├── brats2d_sag/
│   ├── train_gli/                   
│   ├── tran_met/                    
│   ├── train_met_add/               
│   ├── val_gli/             
│   └── val_met/  
└── brats3d/
    ├── train_gli/                  
    ├── tran_met/                 
    ├── train_met_add/              
    ├── val_gli/              
    └── val_met/ 

3. Validation

To get the segmentation scores on the official validation dataset, you will have to obtain segmentation pseudo-labels first. You can reuse the ones from the data preparation step. Put the obtained segmentations into /work_path/brats3d/pseudo_random/gli_segmentations and /work_path/brats3d/pseudo_random/gli_segmentations. Remember to replace /work_path with your actual work directory in the source code. For the segmentation part of the 3d validation phase, since it is from the brats python package, a cuda capable gpu is mandatory.

4. Training

If you bravely want to train the model, the available code uses pytorch and accelerate and using cuda capable gpus is mandatory (since training it on cpu would take ages). Please, check the cuda compute capability of your system before using advanced features such as AMP, model compilation and flash attention.
link --> https://developer.nvidia.com/cuda-gpus

Wandb

Wandb is included in the training script. To make it work, make sure to log it into wandb from you terminal.

Resources

The model was trained on 2/4 NVIDIA L40S (48GB each). Adjust your settings and your batch size accordingly.

5 Checkpoints

You can find the best checkpoints at: https://drive.google.com/drive/folders/1u7mJN9BZ-a5HLz8IMDrU0T1P3UEYGGe7?usp=sharing Initialize your model like: model = HFGAN(dim=64, num_inputs=4, num_outputs=1, dim_mults=(1,2,4,8,10), n_layers=4, skip=True, blocks=False,grouped_encoder=grouped_encoder, infuse_view=False) Where: ''grouped_encoder == True''' if you want to run a checkpoint from run24, '''False''' otherwhise.

6. Issues?

Feel free to contact me at: 269868@studenti.unimore.it if you got any issues running the code.