G-UNETR++: A Gradient-Enhanced Network for Accurate and Robust Liver Segmentation from CT images

---

🔥G-UNETR++: A Gradient-Enhanced Network for Accurate and Robust Liver Segmentation from CT images
Paper: G-UNETR++

---

Requirements

Our code is based on UNETR++ code.
But, we modified the code for easy implementation. Our GPU is RTX 3090 GPU.

Environment

1. Create and activate conda environment

conda create --name gunetr_pp python=3.9
conda activate gunetr_pp

2. Install pytorch

# cuda 11.3
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.3 -c pytorch

It is important that check your cuda version.
Please, see the pytorch document.

3. Install other dependencies

pip install -r requirements.txt

---

Dataset

In paper, we teseted LiTS, 3Dircadb, and Sliver07.

Dataset format

GUNETR_pplus_LiTS
├── DATASET_Synapse                  
│   ├── unetr_pp_raw
│       ├── unetr_pp_raw_data           
│           ├── Task02_Synapse           
│               ├── Task002_Synapse         
│                   ├── seg_gt
│                       ├── 3Dircadb
│                       ├── LiTS
│                           ├── segmentation-3.nii
│                           ├── segmentation-5.nii
│                           ├── ...
│                           └── segmentation-127.nii
│                       ├── Sliver07
│                   ├── unetr_pp_Data_plans_v2.1_stage1
│                       ├── 3Dircadb
│                       ├── LiTS
│                           ├── volume-3.nii
│                           ├── volume-5.nii
│                           ├── ...
│                           └── volume-127.nii
│                       ├── Sliver07
│                   └── unetr_pp_Plansv2.1_plans_3D.pkl

LiTS dataset: 131 cases.
3Dircadb link: 20 cases.
Sliver07 link: 20 cases.

Our LiTS-testset number is 3, 5, 15, 18, 28, 33, 37, 42, 47, 54, 62, 70, 73, 80, 90, 100, 105, 110, 121, and 127.

Model Checkpoint

GUNETR_pplus_LiTS
├── output_synapse                 
│   ├── 3d_fullres
│       ├── Task002_Synapse                   
│           ├── unetr_pp_trainer_synapse__unetr_pp_Plansv2.1        
│               ├── fold_4
│                   ├── validation_raw
│                   ├── model_best.model
│                   └── model_best.model.pkl

Best-model-chekcpoint: link.

---

Implementation

1. Make npy files

$> python LiTS_npy_make.py

You select the options, LiTS, 3Dircadb, and Sliver.

2. Evaluation script

$> cd ./evaluation_scripts
$> sh run_evaluation_synapse.sh

You select the options, LiTS, 3Dircadb, and Sliver.

3. Calculation metrics Please see our jupyter notebook.
   We implemented all of metric classes.

You can control post-processing option through flag_post = True.

---

Result

LiTS

Model	DSC	Jaccard	VOE	RAVD	ASSD	RMSD	MSSD
Guo et al.	0.9430	---	---	---	2.30	4.70	34.70
Song et al.	0.9680	---	0.0700	0.0150	---	---	---
Lei et al.	0.9630	---	0.0688	0.0146	1.37	77.60	---
Chen et al.	0.9650	---	0.0670	0.0090	1.22	28.09	---
Zhu et al.	0.9688	0.9422	0.0578	0.0039	1.09	---	16.08
Chen et al.	0.9727	---	0.0531	1.0800	1.31	3.05	---
Ours (G-UNTER++)	0.9737	0.9490	0.0511	0.0201	0.64	1.17	12.75

---

References

UNETR++

---

Citation

@ARTICLE{
  title={G-UNETR++: A gradient-enhanced network for accurate and robust liver segmentation from CT images}, 
  author={Seungyoo Lee, Kyujin Han, Hangyeul Shin, Harin Park, Xiaopeng Yang, Jae Do Yang, Hee Chul Yu, Heecheon You},
  journal={}, 
  year={2024},
  doi={}}