TriALS: Triphasic-Aided Liver Lesion Segmentation Benchmark

Welcome to the official repository for the TriALS challenge at MICCAI 2024. This repository is built on nnUNet, incorporating additional functionalities for training models such as SAMed ✂️, Mamba Variants :snake:, and MedNeXT ⚡. We aim to establish a benchmark for MICCAI 2024 challenges.

News & Updates

• 2024/8/8: TriALS is now open for submissions! Find detailed instructions here.
• 2024/4/4: Pre-registration for the TriALS challenge is now open Pre-register here.
• 2024/4/4: The TriALS repository now supports SAMed ✂️, Mamba Variants :snake:, and MedNeXT ⚡.

Models Available

• nnU-Net V2 ✓
• Swin-UNETR ✓
• SegResNet ✓
• MedNeXt ✓
• LightM-UNet ✓
• U-Mamba ✓
• SAMed ✓
• SAM-3D (In Progress)

Benchmark Available

• Liver Tumor Segmentation 5-Fold (100 epoch) (LiTS) ✓

The Dice coefficient for lesions is calculated as an average across all lesions. For the evaluation code and a detailed list of metrics, please refer to the evaluation folder.

• Liver Tumor Segmentation 5-Fold (1000 epoch) (LiTS) (in-progress)

• TriALS (in-progress)

Getting Started

To get started, follow these steps:

1. Clone the Repository

   git clone https://github.com/xmed-lab/TriALS
   

2. Create and Activate a Virtual Environment

   conda create -n TriALS python=3.10
   conda activate TriALS
   

3. Install Pytorch: Follow the instructions here:

    pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
   

4. Install the Repository

    cd TriALS
    pip install -e .
   

Data Preparation

We follow the nnU-Net V2 guideline for data preparation, detailed below and accessible here.

1. Download and Prepare the MSD Liver Dataset, and extract it into the data/nnUNet_raw_data_base directory.

   gdown https://drive.google.com/uc?id=1jyVGUGyxKBXV6_9ivuZapQS8eUJXCIpu
   # or use wget
   wget https://msd-for-monai.s3-us-west-2.amazonaws.com/Task03_Liver.tar
   

   tar -xvf Task03_Liver.tar -C data/nnUNet_raw_data_base
   

2. Export nnUNet-v2 evironment variables:

   export nnUNet_raw=<path-to>/data/nnUNet_raw_data_base/
   export nnUNet_preprocessed=<path-to>/data/nnUNet_preprocessed/
   export nnUNet_results=<path-to>/data/nnUNet_results/
   

3. Convert the MSD Liver dataset to nnU-Net format:

    python nnunetv2/dataset_conversion/convert_MSD_dataset.py -i data/nnUNet_raw_data_base/Task03_Liver/
   

Sanity Check: Please verify that the dataset is organized in the following structure:

data/
├── nnUNet_raw_data_base/
│   ├── Dataset003_Liver/
│   │   ├── imagesTr
│   │   │   ├── liver_1_0000.nii.gz
│   │   │   ├── liver_2_0000.nii.gz
│   │   │   ├── ...
│   │   ├── labelsTr
│   │   │   ├── liver_1.nii.gz
│   │   │   ├── liver_2.nii.gz
│   │   │   ├── ...
│   │   ├── dataset.json

3. Preprocess the LiTS Dataset: Replace <DATASET_ID> in the command below with 3:

   nnUNetv2_plan_and_preprocess -d <DATASET_ID> --verify_dataset_integrity
   

Model Training

To train the models, follow these instructions:

• General Training Command on GPU 0. For the preliminary benchmark, all models are trained for 100 epochs. Please note that all model variants, with the exception of nnUNet, are trained without deep supervision.

  CUDA_VISIBLE_DEVICES=0 nnUNetv2_train <DATASET_ID> <CONFIGURATION> <FOLD_NUM> -tr <TRAINER>
  

  Model	Configuration	Trainer
  nnU-Net 2D	2d	nnUNetTrainer_100epochs
  nnU-Net 3D	3d_fullres	nnUNetTrainer_100epochs
  Swin-UNETR	3d_fullres	nnUNetTrainerSwinUNETR_100epochs
  SegResNet	3d_fullres	nnUNetTrainerSegResNet_100epochs
  U-Mamba	3d_fullres	nnUNetTrainerUMambaBot_100epochs
  LightM-UNet	3d_fullres	nnUNetTrainerLightMUNet_100epochs
  MedNext-B-5	3d_fullres	nnUNetTrainerV2_MedNeXt_B_kernel5_100epochs
  MedNext-L-5	3d_fullres	nnUNetTrainerV2_MedNeXt_L_kernel5_100epochs
  SAMed-B	2d_p256	nnUNetTrainerV2_SAMed_b_r_4_100epochs
  SAMed-H	2d_p512	nnUNetTrainerV2_SAMed_h_r_4_100epochs

Example Training Commands

• U-Mamba :snake:

CUDA_VISIBLE_DEVICES=0 nnUNetv2_train 3 3d_fullres 0 -tr nnUNetTrainerUMambaBot

• MedNeXT ⚡

CUDA_VISIBLE_DEVICES=0 nnUNetv2_train 3 3d_fullres 0 -tr nnUNetTrainerV2_MedNeXt_B_kernel5

• SAMed ✂️

Download the checkpoint of original SAM into checkpoint

# sam-b checkpoint
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth -O checkpoints/sam_vit_b_01ec64.pth
# sam-h checkpoint
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth -O checkpoints/sam_vit_h_4b8939.pth

CUDA_VISIBLE_DEVICES=0 nnUNetv2_train 3 2d_p256 0 -tr nnUNetTrainerV2_SAMed_b_r_4

Inference

• Validation Inference

To generate model validation outputs, execute the command below. To acquire probabilities for the optimal configuration, append the --npz flag, noting this requires significant disk space.

CUDA_VISIBLE_DEVICES=0 nnUNetv2_train DATASET_ID CONFIGURATION FOLD_NUM -tr TRAINER --val --npz

• Inference on unseen

nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d DATASET_ID -c CONFIGURATION -tr TRAINER_NAME

• Best Configuration and Ensembles (in progress)

Citation

If you utilize the TriALS dataset, please cite the relevant paper:

@misc{elbatel2026trialstriphasicaidedliverlesion,
      title={TriALS: Triphasic-Aided Liver Lesion Segmentation Benchmark in Non-Contrast CT}, 
      author={Marawan Elbatel and Mohamed Ghonim and Jiaji Mao and Zhuosheng Lin and Katharina Eckstein and Andrés Martínez Mora and Jonathan Deissler and Maximilian Rokuss and Constantin Ulrich and Zdravko Marinov and Wenhui Deng and Baoxun Li and Huijun Hu and Jun Shen and Mohanad Ghonim and Khadiga Omar Nassar and Mariam Elbakry and Menna Dyab and Amr Muhammad Abdo Salem and Nouran Elghitany and Noha Elghitany and Yi Qin and Xuanqi Huang and Haonan Wang and Shao-Woo Yen and Ahmed Elghamry Saba and Salma Ahmad and Xinyan Fang and Jiahao Zhang and Xiaodi Wang and Xinghua Ma and Gongning Luo and Jessica C. Delmoral and João Manuel R. S. Tavares and Ankan Deria and Adinath Dukre and Yutong Xie and Imran Razzak and Dongwook Kim and Matthew Choi and Hanxiao Zhang and Minghui Zhang and Xin You and Abdul Qayyum and Steven A. Niederer and Moona Mazher and Rachika E. Hamadache and Ricardo Montoya-del-Angel and Robert Martí and Xavier Lladó and Toufiq Musah and Livingstone Eli Ayivor and Enrique Almar-Munoz and Agnes Mayr and Kaouther Mouheb and Esther E. Bron and Stefan Klein and Ahmed Abouelhoda and Amira Adel and Susan Adil Ali and Rainer Stiefelhagen and Klaus H. Maier-Hein and Fabian Isensee and Aya Yassin and Xiaomeng Li},
      year={2026},
      eprint={2605.16572},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.16572}, 
}

If you utilize the baselines in this repository for your research, please consider citing the relevant papers for Swin-UNETR, SegResNet, LightM-UNet, U-Mamba, MedNext, SAMed, and nnU-Net.

Acknowledgements

We would like to acknowledge the contributions of nnU-Net and the authors of the baseline models: LightM-UNet, MedNeXT, and SAMed. This repository builds upon their foundational code and work.