MSWAL : 3D Multi-class Segmentation of Whole Abdominal Lesions Dataset (MICCAI2025 accepted)

📖 Paper on arxiv: https://arxiv.org/pdf/2503.13560

💽 Dataset on huggingface(only training dataset has been released): https://huggingface.co/datasets/zhaodongwu/MSWAL/tree/main

This repo presents the implementation of the MSWAL 🩻

Key Features of MSWAL 🔑

• Comprehensive Coverage: MSWAL includes seven common abdominal lesions (gallstones, kidney stones, liver tumors, kidney tumors, pancreatic cancer, liver cysts, and kidney cysts), addressing the single-organ focus of existing datasets.

• Fine-Grained Annotation: Lesions are specifically categorized by type, unlike "pan-cancer" labeling, enabling precise diagnostic suggestions and reducing radiologists' workload.

• Full Annotation: MSWAL is fully annotated with no missing labels, minimizing noise and enhancing model performance compared to partially annotated datasets.

Stastics of MSWAL 📊

MSWAL comprises 694 high-resolution CT scans (191,417 slices) with 7 lesion types (e.g., liver tumors, kidney stones), balanced gender distribution (53.9% male, 46.1% female), and diverse CT contrast phases (non-contrast, arterial, venous, etc.), offering a fully annotated, multi-class dataset for precise abdominal lesion segmentation.

Examples of MSWAL 📊

Inception nnU-Net structure 🧩

Get Started ⚡️

Inception nnU-Net in this paper is highly dependent on the preprocessing and architecture of nnU-Net. You can find how nnU-Net works here: nnU-Net

Step 1: Requirements.

Conda create -n IncUnet --python == 3.9
git clone https://github.com/MIC-DKFZ/nnUNet.git
cd nnUNet
pip install -e .

Step 2: Preprocess the MSWAL.

Create a folder for MSWAL, which has three sub-folders: nnU-Net_raw; nnUNet_preprocessed; nnUNet_results. More details can be seen in nnU-Net's dataset_format.

MSWAL Folder Structure

Dataset/
├── nnU-Net_raw/
│ ├── dataset.json # Dataset configuration file
│ ├── imagesTr/ # Training images
│ ├── imagesTs/ # Testing images
│ ├── labelsTr/ # Training labels
│ └── labelsTs/ # Testing labels
├── nnUNet_preprocessed/ # Preprocessed data (automatically generated by nnU-Net)
└── nnUNet_results/ # Training results and model outputs

Instructions:

1. Save the dataset in nnU-Net_raw following the nnU-Net dataset format.
2. Modify environment variables to point to the MSWAL folder for nnU-Net processing.

vim ~/.bashrc

Set the path to the dataset if you are using Linux.

export nnUNet_raw='data_fold_path/nnUNet_raw'
export nnUNet_preprocessed='data_fold_path/nnUNet_preprocessed'
export nnUNet_results='data_fold_path/nnUNet_results'

Then update the environment variables.

source ~/.bashrc

If you want to reproduce the Inception nnU-Net on MSWAL, please modify epoch to 1500 and initial learning rate to 0.001 in nnUNetTrainer.py(OPTIONAL). Now you can preprocess the dataset.

nnUNetv2_plan_and_preprocess -d DATASET_ID --verify_dataset_integrity -c 3d_fullres -p nnUNetResEncUNetLPlans

Step 3: install dynamic_network_architectures

git clone https://anonymous.4open.science/r/MSWAL--406B.git
cd dynamic_network_architectures
pip install -e .

Step 4: train the model.

Modify network_class_name in 3d_fullres to "dynamic_network_architectures.architectures.inception.InceptionNnunet" in nnUNetResEncUNetLPlans.json and change the relevant batch_size to smaller (dependent on your GPU). Then run the training process. We use five-fold cross-validation to train the model.

nnUNetv2_train DATASET_ID 3d_fullres 0 -p nnUNetResEncUNetLPlans
nnUNetv2_train DATASET_ID 3d_fullres 1 -p nnUNetResEncUNetLPlans
nnUNetv2_train DATASET_ID 3d_fullres 2 -p nnUNetResEncUNetLPlans
nnUNetv2_train DATASET_ID 3d_fullres 3 -p nnUNetResEncUNetLPlans
nnUNetv2_train DATASET_ID 3d_fullres 4 -p nnUNetResEncUNetLPlans

Step 5: run inference

nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d DATASET_NAME_OR_ID -c 3d_fullres -f 0 1 2 3 4 -p nnUNetResEncUNetLPlans

Step 6: evaluation

nnUNetv2_evaluate_folder -djfile path/dataset.json -pfile path/plans.json path/labelsTs path/infersTs