Rapid automatic fat quantification in thigh muscle dystrophy for muscular dystrophies [Huashan Neuromuscular Group]

The toolkit was developed for rapid and automatic fat quantification for thigh muscle degeneration in muscular dystrophy patients.

This project takes advantage of nnU-Net (a self-adapting framework for U-Net-based medical image segmentation) and MONAI (a PyTorch-based framework for deep learning in healthcare imaging) to build the model. The training dataset included quantitative MRIs (IDEAL) from 33 muscular dystrophy patients and 17 healthy controls.

1. Download the whole repository to local

git clone https://github.com/Hirriririir/Rapid-automatic-fat-quantification-for-thigh-muscle-degeneration

OR directly download the whole repository from this link.

2. Install required dependencies

# for nnU-Net pipeline
pip install nnunet

# for calculating fat fraction
pip install monai

3. Place your MRI file in the right place (Input_data)

Input_data: Two kinds of IDEAL (3-point Dixon) sequences, namely fat and water, are required. You can rename your files according to a format like [Thigh_001_0000.nii.gz], where '001' represents subject number, and '0000' and '0001' represent 'Fat IDEAL sequence' and 'Water IDEAL sequence' respectively.
Output_segmentation: The segmentation of the mask will be generated after running a inference process by a pretrained nnU-Net model.
Output_fat_fraction: A csv. file recording the fat fractions of all thigh muscles in each subject will be generated after running corresponding code.
nnUNet_Results_Folder: Download pretrained model "model_final_checkpoint.model" to the right place "./nnUNet_Results_Folder/nnUNet/3d_fullres/Task501_ThighMuscles/nnUNetTrainerV2__nnUNetPlansv2/fold_0"

├── Input_data
│   ├── Thigh_001_0000.nii.gz
│   ├── Thigh_001_0001.nii.gz
│   ├── Thigh_002_0000.nii.gz
│   ├── Thigh_002_0001.nii.gz
│   ├── Thigh_003_0000.nii.gz
│   └── Thigh_003_0001.nii.gz
├── Output_fat_fraction
├── Output_segmentation
│   ├── Thigh_001.nii.gz
│   ├── Thigh_002.nii.gz
│   ├── Thigh_003.nii.gz
│   └── plans.pkl
├── Segmentation and fat fraction .ipynb
├── nnUNet_Results_Folder
│   └── nnUNet
│       └── 3d_fullres
│           └── Task501_ThighMuscles
│               └── nnUNetTrainerV2__nnUNetPlansv2.1
│                   ├── fold_0
│                   │   ├── debug.json
│                   │   ├── model_final_checkpoint.model
│                   │   ├── model_final_checkpoint.model.pkl
│                   │   ├── postprocessing.json
│                   └── plans.pkl
├── nnUNet_preprocessed
└── nnUNet_raw_data_base
    ├── nnUNet_cropped_data
    └── nnUNet_raw_data

4. Run the code in the Jupyter notebook

# Inference using Our pre-trained model in the 'nnUNetTrainerV2__nnUNetPlansv2.1' dictionary

!nnUNet_predict -i './Input_data' -o './Output_segmentation' -t 501 -tr nnUNetTrainerV2 -m 3d_fullres

5. Results

Segmentation results

These MRIs are from the Test dataset that the model had not seen during the training process. Our model shows a good segmenting ability compared to a human radiologist.

We recommend using ITK-SNAP to visualize the MRI files, and the label descriptions for this project can be found as 'Label descriptions for ITK-SNAP.label' in this repository.

Fat fraction visualization

Fat fraction results

Subject	SA	RF	VL	VI	VM	AM	GR	BL	ST	SM	BB
Thigh_001	0.182836	0.537026	0.309223	0.343359	0.300494	0.648822	0.235308	0.639768	0.649394	0.617633	0.418268
Thigh_002	0.163060	0.501802	0.536133	0.529328	0.503416	0.405981	0.118544	0.618817	0.542495	0.672286	0.156817
Thigh_003	0.205246	0.215608	0.330503	0.369018	0.349797	0.224268	0.127358	0.325287	0.201906	0.266562	0.198464

BL, Biceps femoris long head; BB, Biceps femoris short head; ST, Semitendinosus; SM, Semimembranosus; AM, adductor magnus; VI, Vastus intermedius; VL, Vastus lateralis; VM, Vastus medialis; RF, Rectus femoris; GR, Gracilis; SA, Sartorius.

Cpresident/Rapid-automatic-fat-quantification-for-thigh-muscle-degeneration