Awesome
GFE-Mamba
This code is the pytorch implementation of the GFE-Mamba which has been accepted by XX named "GFE-Mamba: Mamba-based AD Multi-modal Progression Assessment via Generative Feature Extraction from MCI". It can be used to accurately predict whether a patient will process to AD from MCI in a period (symbolled as $\Delta t$) with Multimodal (including MRI, PET and multiple scale assessments) information without the PET engagement.
<br><br>
<br><br>
Metrics of AD prediction on SOTA models on 1-year Dataset
| Method | Precision | Recall | F1-score | Accuracy | MCC |
|---|---|---|---|---|---|
| Resnet50 | 81.03% | 78.95% | 68.18% | 73.17% | 59.00% |
| Resnet101 | 79.31% | 47.37% | 81.82% | 60.00% | 50.57% |
| TabTransformer | 84.29% | 93.33% | 75.68% | 83.58% | 70.22% |
| XGBoost | 88.14% | 86.53% | 86.92% | 87.37% | 73.90% |
| Qiu et al's | 86.67% | 81.82% | 81.82% | 81.82% | 71.29% |
| Fusion model | 89.83% | 88.21% | 88.91% | 89.87% | 78.06% |
| Zhang et al's | 76.67% | 83.33% | 45.45% | 58.82% | 48.42% |
| GFE-GAN(Ours) | 95.71% | 93.33% | 96.55% | 94.92% | 91.25% |
Pre-requisties
- Linux
- python==3.10
- NVIDIA GPU (memory>=14G) + CUDA cuDNN
Python libraries
- accelerate == 0.27.2
- tqdm == 4.66.2
- monai == 1.3.0
- monai-generative == 0.2.3
- torch == 2.0.0+cu118
- torchvision == 0.15.0+cu118
- matplotlib == 3.8.0
- PyYAML == 6.0.1
- einops == 0.7.0
- torchmetrics == 1.3.1
- torch-fidelity == 0.3.0
How to train
Preparation of the data
The method of preparation of two datasets are included in the sub-directory: "./GEF-Mamba_ADNI_Dataset". Please refer to README there.
MRI to PET Generation training
First, accordding to the dataloader, the MRI and PET should both be "nii.gz" files and should be put into one subdirectory. Then MRI and PET should be named as "MRI.nii.gz", "PET.nii.gz", repectively.
|--MRI-PET
|----patient1
|------MRI.nii.gz
|------PET.nii.gz
|----patient2
|------MRI.nii.gz
|------PET.nii.gz
|----patient3
|------MRI.nii.gz
|------PET.nii.gz
...
After the configuration of directory, to make sure the "train_path" and "eval_path" have been specified. Then run the command.
python main_gan_vit.py
Prediction training
For this part, a MRI to PET generation network should be trained, a pre-trained 3D GAN-Vit parameter could be downloaded and used. We assume that we have the table containing temporal information of whether one patient was diagnosed as MCI or AD. Since the program will read the temporal information and make a mapping to images, the name of MRI should be:
[patient_name]-[diagnosis_time]-[label]
For example if a patient called 941_S_4187 is diagnosed at 2012_02_29_14_05_50.0 will not become AD next diagnosis, then this MRI is named as:
941_S_4187-2012_02_29_14_05_50.0-0.nii.gz
Then run the command
python classify_mamba.py
Model Evaluation
Download the pre-trained 3D GAN-Vit and Mamba Classifier parameters first and specify the yaml configuration file.:
python test_mamba.py
Citation
@article{fang2024gfe,
title={GFE-Mamba: Mamba-based AD Multi-modal Progression Assessment via Generative Feature Extraction from MCI},
author={Fang, Zhaojie and Zhu, Shenghao and Chen, Yifei and Zou, Binfeng and Jia, Fan and Qiu, Linwei and Liu, Chang and Huang, Yiyu and Feng, Xiang and Qin, Feiwei and others},
journal={arXiv preprint arXiv:2407.15719},
year={2024}
}