https://doi.org/10.5061/dryad.ngf1vhj1m

The following figure depicts our Individualized real-time functional MRI closed-loop neuromodulation (iNM) Intervention.

A. iNM strengthens visual perception and visual imagery networks: 1) and 2) High-resolution anatomical images were acquired and registered to the Siemens’ console computer; 3) Each participant’s individualized MT and MST networks were delineated and contoured; 4) iNM data extracted from individualized networks was preprocessed and general linear modeling was used to decode each coherence level in real-time every TR=2 sec (yellow line denotes the time series during green, purple, and black periods representing up or down direction at coherence level C100, C33, and baseline-random motion period, respectively); 5) The BOLD signal intensity and spatial extent (individualized network) for each direction was computed via a GLM, and beta weights were updated, see Fig. 1B. 6) iNM interface shows the extent of the circle filled, which directly corresponds to the percent upregulation (red fill) or downregulation (blue fill) of the BOLD signal intensity and spatial extent.
B. Task design: The study was performed over two days in which 5 control and 5 iNM runs were completed in an alternating fashion. Each run was performed in a temporal sequence that included: a coherent motion block lasting 20 seconds interleaved with a 10-second baseline-random motion block. The level of coherence was counterbalanced across runs, and blocks ensuring that the same coherence levels were not stacked back-to-back.
C. Computation of the iNM signal: The neuromodulation stimulus was calculated as a percent change of the BOLD signal intensity during the task compared to the baseline-random motion block during each control run which served as a reference. If the percent BOLD signal change during a 2-sec interval was higher than the 10th percentile, the iNM interface indicated a 10% upregulation, and if less than the 10th percentile, the iNM interface indicated downregulation.
D. Brain activation maps show increased activation in the iNM condition compared to control: GLM-generated activation maps for each coherence level and motion direction in iNM (top) and control- no cNMT (bottom) conditions are shown; voxelwise p-value = 0.05, cluster size > 20 voxels.

Description of the data and file structure

In the root directory (Data), each participant's data is organized into individual folders. These participant folders contain subfolders that accommodate specific aspects of the study. Notably, instances where participants repeated the task in a different visual field quadrant are indicated by the prefix TT_Q in the subfolder names. Within each quadrant's folder, there are two subfolders labeled down and up, corresponding to the task direction. Each of these subfolders contains two additional folders named control and iNM, which correspond to the control and neuromodulation arms of the study. Due to HIPPA compliance regarding the sharing of MRI data that could potentially be reconstructed into facial images, and because certain steps in the AFNI preprocessing pipeline require the presence of the skull, we have shared only the preprocessed files. The preprocessing code we used is included in the repository for your reference.

Data/
 └──HS002/
  ├── ...
 └──HS003/
  └── TT_Q1/
   └── preprocessedTT27/
    └── up                                                                                                                                                                                                                                                                 └── control/
          ├── ...                                                                                                                                                                                                                                                         └── iNM/
          ├── ...               
    └── down/                              
                   └── control/
          ├── ...                                                                                                                                                                                                                                                         └── iNM/
          ├── ...   
  └── TT_Q3/
   └── ...
 └──HS004/
  ├── ...
 ...
 └──HS013/
  ├── ...

Code Overview

Before any code is run, AFNI must be installed on the computer. Instructions to install AFNI can be found here: AFNI Installation Link
In addition, Python must also be installed on the computer. The following python packages are needed for analysis:
pandas == 2.1.3
numpy == 1.26.2

0. Preprocessing
 The script "preprocessing.sh" reads the AFNI Head/BRIK files and preprocesses them. The following steps are performed:
  (1)  Convert dicom to HEAD/BRIK
  (2)  Detection of outliers, despking
  (3)  Slice time correction
  (4)  Alignment of functional volumes (to the first volume of the first run)
  (5)  Skullstripping of anatomical data and Talairaching using template TTN27+tlrc dataset in AFNI
  (6)  Warp functional data to Talairach
  (7)  Blurring with fwhm=6mm
  (8)  Masking
  (9)  Segmentation with 3dSeg
  (10) Scaling to have mean voxel time series of 100

1. Generalized Linear Model (GLM) Labels
 The script "2_generate GLM_labels.py" generates labels for each coherence and direction that will be used to run the GLMs.

2. GLM
 The script "3_run_GLM.py" runs the GLM. We ran 2 GLMs for each coherence level and direction:
  (1) One in which 5 control runs (runs 3-4-5-6-7) were concatenated
  (2) One in which 5 iNM runs (runs 8-9-10-11-12) were concatenated

3. Group Analysis
 The script "4_group_analysis.py" runs a 3DREMLFit on individual subject GLMs to compare for group analysis.

4. Anatomical ROIs
 GLM-generated activation maps have big clusters that span across multiple brain regions. To extract BOLD signal from each brain region, the first step  was to generate anatomical datasets for each region. The script "5_generate_anat_ROIs.py" generates these anatomical masks.
 
5. Intersection Masks
  The script "6_intersection_masks.py" generates masks that encompass the intersection of functional clusters and anatomical regions. These  intersection masks can be obtained using 3dcalc.
  (1) First step is to obtain datasets for functional clusters using 3dClusterize. Functional clusters for iNM condition were based on activation maps for  iNM condition, and functional clusters for control condition were based on activation maps for control condition.
  (2) 3dcalc was then used to multiply anatomical ROIs with the functional clusters just obtained to obtain the intersection masks

6. Find Local Maxima
 The goal was to generate tables of activated regions in each condition. The script "7_extrema_mema.py" runs the command 3dExtrema for each cluster in the activation maps of the corresponding condition to create these tables.
   (1) Functional clusters obtained previously through 3dClusterize were then masked and fed into 3dExtrema to find the local maximum.  Information  regarding the anatomical region in which the functional cluster was located in, the number of voxels of the anatomical region, the number of voxels  that intersected between the anatomical region and the functional cluster, the ratio between the intersection voxels and the total voxels of the  anatomical  region, the X, Y, and Z coordinates of the local maximum and the z-score of the local maximum are stored in an excel file for each cluster.