fmri data_based MDD classification (preprocessing is implemented with DPABI)

上传了系统文档 The system document was uploaded.

将预处理后的功能磁共振数据组织成相应的文件结构，自动实现数据准备、数据分割、模型训练和结果可视化。

Organize pre-processed FMRI data referring to the corresponding file structure, and then the code will automatically realize data preparation, data segmentation, model training, and result visualization

注意在过程中为了防止程序中断需要重新运行程序花费的时间，在运行过程中会自动生成中间文件，这样即使中断了，上一步的处理结果也会以文件的形式保存下来，中间文件包括HC/MDD_splice_along_time, model等，参见Intermediate file

Note that to prevent program interruption, which will lead to re-running the program, during the running process, intermediate files will be automatically generated. In this way, even if the process is interrupted, the results of the previous step will be saved in file form. Intermediate documents include: HC/MDD_splice_along_time, model, etc., see Intermediate file

IF THIS IS USEFUL FOR YOU, STAR FOR ME PLS! THKS!!

Content:

• Usage
  • Input File Format
  • Code Format
  • Code Usage
  • Intermediate file
• Rationale

• Notice

• Ref-paper

Usage:

基于fmri数据的ml&dl方法疾病分类

Classification of diseases using ML and DL methods based on FMRI data

通过将预处理后的数据组织到相应的文件中，自动实现数据的划分、模型的训练和模型的可视化绘制

By organizing the pre-processed data into the corresponding files, the training of the data division model and the visualization of the model are realized automatically

Input File Format：

对原始fMRI数据用DPABI工具进行处理后（包括脑区分割）；

Raw fMRI data processed with the DPABI tool (addition with brain segmentation)

用GRENTA工具进行SFC (static functional connection)和DFC (dynamic functional connection)的提取；

GRENTA tool was used to extract SFC (static functional connection) and DFC (dynamic functional connection) features.

得到每个被试的SFC和DFC矩阵,为index.mat文件格式，最终输入文件格式应该如下：(SFC中的mat文件是2维的，DFC中的mat文件是3维的，有一个维度是时间)

The SFC and DFC matrix of each subject is obtained in the file format index.mat. The final input file format should be as follows: (mat file in SFC is 2-dimensional; the mat file in DFC is 3-dimensional, with one dimension of time.)

Code Format:

• anatomical.py

做特征选择之后，对选择出来的特征进行生理解释，也就是进行区域的对应，找出选择出来的特征是哪些区域之间的功能连接)

After feature selection, physiological interpretation is carried out on the selected features, that is, regions related to the selected functional connection feature will be printed out)

• draw.py

训练和测试结束之后，绘制gridsearch的过程，绘制AUC曲线

After training and testing, draw the grid-search process and draw the AUC

• svm.py

分类模型pipeline

Classification model pipeline

• main.py

主函数，输入HC和MDD文件的路径，FC类别，模型种类(SVM/LSTM);
进行自动特征处理，t test特征选择，SVM-RFE特征选择，SVM分类, Gridsearch寻找最佳参数，以及特征的生理解释的自动对应输出，测试结果的AUC图像的自动绘制。

The main function, Input: HC and MDD file path, SFC/DFC, model class (SVM/LSTM); Perform automatic feature processing, t-test feature selection, SVM-RFE feature selection, SVM classification, Grid search to find the best hyperparameters, as well as automatic print physiological interpretation of features (regions of the brain), automatic rendering of AUC images of test results

Code Usage:

python main.py SVM sfc "SFC_Data\HC\HC_Data" "SFC_Data\MDD\MDD_Data" --Threshold 0.2 --atlas AAL

Intermediate file

参考Rationale理解

• HC/MDD_splice_alone_time

一类人的SFC/DFC矩阵拉直之后堆叠成的矩阵

• SFC

直接拉伸成一维，所有人的堆叠，得到二维矩阵

• DFC

对每一个人，将DFC的每一个矩阵拉伸成一维，将所有的一维向量拼接在一起；接着将所有人的拉伸拼接后的向量堆叠在一起，得到一个矩阵，一个维度是人的编号，一个维度是FC特征

• HC/MDD_stack_alone_time

一类人的DFC矩阵，每一个人的DFC中的每一个FC拉伸之后，将他们进行堆叠成一个二维矩阵，一个维度是时间；接着将所有人的矩阵堆叠在一起得到一个三维矩阵，一个维度是人编号，一个维度是时间，还有一个维度是拉伸后的FC

• model

存储训练后的模型，之后在进行生理解释和可视化的时候，可以直接获取训练好的模型进行后续运算

• SFC

SFC相关结果图

• DFC

DFC相关结果图

Rationale：

• SFC数据准备
  
• DFC数据准备
  
• 分类器
  
• 选择特征的生理解释
  
  
• 选择特征的可视化
  用BrainNet Viewer工具进行选择出来的特征的可视化
• 代码框架
  

Notice：

1. main.py中实现了多进程，用了15核的CPU，在运行之前请将其更改为合适的个数，代码在

p = Pool(processes=15)
# 把15改成你的计算机的cpu内核数

1. LSTM方法的实现还未上传
2. 输入文件目录结构一定要按照上文描述组织

Ref_paper:

• Paper_ref[^1]

[^1]: Castellazzi, Gloria, et al. “A machine learning approach for the differential diagnosis of Alzheimer and Vascular Dementia Fed by MRI selected features.” Frontiers in neuroinformatics 14 (2020): 25.