🏆 We Won the Champion of MICCAI FLARE2024 Challenge Task1!

🔗 Key Links

• FLARE24 Official Website: Task1

• Docker Download: Baidu NetDisk or Google Drive

  You only need to place all the .nii.gz files ready for inference in the inputs directory. Then, run the following commands to start inference:

  docker load -i gmai.tar.gz
  docker container run --gpus "device=all" --name gmai --rm \
    -v $PWD/inputs/:/workspace/inputs/ \
    -v $PWD/outputs/:/workspace/outputs/ \
    gmai:latest /bin/bash -c "sh predict.sh"

• Slides Download: View our technical presentation and findings Baidu NetDisk or Google Drive

• 中文介绍：知乎

🔍 A 5.6M Parameter U-Net for Efficient Whole-body Tumor Segmentation

🆚 Key Differences Between Our Method and the default nnU-Net

Configuration	Our Method 🚀	nnUNetV2 🗿
Number of Stages ⬇️	4	6
Training Epochs ⏱️	2000	1000
Learning Rate 📊	0.001	0.01
Target Spacing 📐	2, 2, 3	0.8, 0.8, 3

💡 Our method achieves efficient segmentation with fewer stages and lower resample resolution.

📂 Key Files in this Repository

Only two files you need to focus on:

1. nnUNetTrainer_varianst.py

   • Custom nnUNet trainer with modified training parameters

2. plans.json

   • Network architecture and configuration settings

How to Reimplement?

🛠️ Environment Setup

First, ensure you have PyTorch > 2.0 installed with CUDA support. We conducted our experiments using nnUNet v2.2, which might slightly differ from the latest version nnUNet v2.4. Set up your environment by running:

conda create -n FLARE24_gmai
conda activate FLARE24_gmai
pip install -e .

📂 Data Preparation

• 📁 Dataset: FLARE24 Task1
• 🔢 Data Usage: 5000 partial labeled data (treated as fully labeled without any special handling)
• ⚙️ Preprocessing: Default nnUNet procedure

Organize your labeled data in nnUNet_raw in the following structure:

Dataset024_FLARE24_Task1/
├── imagesTr/
│   ├── coronacases_001_0000.nii.gz
│   ├── ...
│   └── (all 5,000 labeled images ending with .nii.gz)
├── labelsTr/
│   ├── coronacases_001.nii.gz
│   ├── ...
└── dataset.json

You can find the dataset.json file here

📝 Data Preprocessing

1. 🧬 Extract Fingerprints and Plan the Experiment:

nnUNetv2_extract_fingerprint -d 24
nnUNetv2_plan_experiment -d 24

2. 🛠️ Modify the Plans:

Edit the plans file in your nnUNet_preprocessed directory. Refer to our plans.json for guidance. We modified the "patch_size" and "spacing" under "3d_fullres" and create a new configuration "3d_fullres_S4D2W32".

3. ⚙️ Preprocess the Data:

nnUNetv2_preprocess -d 24 -c 3d_fullres -np 4

🚀 Training and Inference

🔧 Training the Model:

Train the network using the following command:

nnUNetv2_train 24 3d_fullres_S4D2W32 all -tr nnUNetTrainer_Epoch2000_Lr1e3

Alternatively, you can use our pre-trained model by copying it into your nnUNet_results folder. You can find our trained model here

🔍 Inference:

To perform inference, run:

nnUNetv2_predict -i ./inputs -o ./outputs -c 3d_fullres_S4D2W32 -f all -d 24 -tr nnUNetTrainer_Epoch2000_Lr1e3

Results

📊 Quantitative Results

Methods	Public Validation		Online Validation		Testing
	DSC(%)	NSD(%)	DSC(%)	NSD(%)	DSC(%)	NSD(%)
Ours	25.34 ± 31.56	24.40 ± 27.80	-	-	-	-

🖼️ Visualization

Two examples with good segmentation results and two examples with bad segmentation results in the validation set.