We will briefly introduce the fundamental concepts of NeuroGraph through self-contained examples. We closely follow the data representation format of PyG. Therefore, interested readers are referred to the PyG documentation for an introduction to the graph machine learning and PyG's data representation formats.
NeuroGraph provides two classes for loading static and dynamic benchmark datastes.
NeuroGraph utilizes the PyG InMemoryDataset class to facilitate the loading of datasets. this allows an easy-to-use interface for applying graph machine learning pipelines. For example, the HCPGender benchmark can be loaded as follows:
from NeuroGraph.datasets import NeuroGraphDataset
dataset = NeuroGraphDataset(root="data/", name= "HCPGender")
print(dataset.num_classes)
print(dataset.num_features)To efficiently store and utilize the dynamic datasets in PyG Batch format, we provide the corresponding functionality. Here is an example of loading the `DynHCPGender dataset:
from NeuroGraph.datasets import NeuroGraphDynamic data_obj = NeuroGraphDynamic(root="data/", name= "DynHCPGender") dataset = data_obj.dataset labels = data_obj.labels print(len(dataset), len(labels))
The dataset is a list of dynamic graphs represented in the PyG batch format, making it compatible with graph machine learning pipelines.
To bridge the gap betwee NeuroGraph and graph machine learning domains, NeuroGraph offers tools to easily preprocess and construct graph-based neuroimaging datasets. Here, we demonstrate how to preprocess your own data to construct functional connectomes and generate corresponding graphs-based representations.
from NeuroGraph import utils fc = utils.preprocess(fmri, regs, n_rois= 1000) # fmri and regs could be numpy arrays
The corresponding Adjacency matrix and PyG data objects can be created from the functional_connectome as follows.
from NeuroGraph import utils
adj = utils.construct_adj(fc, threshold= 5) # construct the adjacency matrix
data = utils.construct_data(fc, label= 1,threshold = 5) # construct PyG data objectWe use correlation as node features while constructing data object from functional connectome.
The following is the source code for processing one fMRI scan with corresponding regressor using our preprocessing pipeline.
from NeuroGraph import utils
import numpy as np
from nilearn.image import load_img
img = load_img("data/raw/1.nii.gz") # 1.nii.gz is fMRI scan
regs = np.loadtxt("data/raw/1.txt") # 1.txt is the movement regressor
fmri = img.get_fdata()
fc = utils.preprocess(fmri, regs, n_rois= 100)
adj = utils.construct_adj(fc, threshold= 5) # construct the adjacency matrix
data = utils.construct_data(fc, label = 1,threshold = 5) # construct torch Data objectOur preprocessing pipeline consists of five steps and can also be applied seperately in steps.
from NeuroGraph import utils
import numpy as np
from nilearn.image import load_img
img = load_img("data/raw/1.nii.gz")
regs = np.loadtxt("data/raw/1.txt")
fmri = img.get_fdata()
parcells = utils.parcellation(fmri,n_rois = 100) ## this uses schaefer atlas by default
Y = utils.remove_drifts(parcells)
Y = utils.regress_head_motions(Y,regs)
fc = utils.construct_corr(Y)
adj = utils.construct_adj(fc, threshold= 5) # construct the adjacency matrix
data = utils.construct_data(fc, label = 1,threshold = 5)NeuroGraph utilizes the HCP1200 dataset as a primary data source for exploring the dataset generation search space and constructing benchmarks. The HCP1200 dataset can be accessed from the HCP website by accepting the data usage terms. Additionally, the dataset is also available on an AWS S3 bucket, which can be accessed once authorization has been obtained from HCP. In this section, we provide various functions that allow you to crawl and preprocess the HCP datasets, enabling the construction of graph-based neuroimaging datasets. These functions streamline the process of obtaining and preparing the data for further analysis and modeling.
from NeuroGraph.preprocess import Brain_Connectome_Rest_Download
import boto3
root = "data/"
name = "HCPGender"
threshold = 5
path_to_data = "data/raw/HCPGender" # store the raw downloaded scans
n_rois = 100
n_jobs = 5 # this script runs in parallel and requires the number of jobs is an input
ACCESS_KEY = '' # your connectomeDB credentials
SECRET_KEY = ''
s3 = boto3.client('s3', aws_access_key_id=ACCESS_KEY, aws_secret_access_key=SECRET_KEY)
# this function requires both HCP_behavioral.csv and ids.pkl files under the root directory. Both files have been provided and can be found under the data directory
rest_dataset = Brain_Connectome_Rest_Download(root,name,n_rois, threshold,path_to_data,n_jobs,s3)The provided function facilitates the download of data from the AWS S3 bucket, performs preprocessing steps, and generates a graph-based dataset. It is important to note that the rest_dataset used in this function consists of four labels: gender, age, working memory, and fluid intelligence. To create separate datasets based on these labels, the following functionalities can be used.
from NeuroGraph import preprocess
rest_dataset = preprocess.Brain_Connectome_Rest_Download(root,name,n_rois, threshold,path_to_data,n_jobs,s3)
gender_dataset = preprocess.Gender_Dataset(root, "HCPGender",rest_dataset)
age_dataset = preprocess.Age_Dataset(root, "HCPAge",rest_dataset)
wm_datast = preprocess.WM_Dataset(root, "HCPWM",rest_dataset)
fi_datast = preprocess.FI_Dataset(root, "HCPFI",rest_dataset)To construct the State dataset, the following functionalities can be used.
from NeuroGraph import preprocess
state_dataset = preprocess.Brain_Connectome_State_Download(root, dataset_name,rois, threshold,path_to_data,n_jobs,s3)If you have the data locally, then the following functionalities can be used to preprocess the data.
from NeuroGraph import preprocess
rest_dataset = preprocess.Brain_Connectome_Rest(root, name, n_rois, threshold, path_to_data, n_jobs)Similarly, for constructing the State dataset, the following function can be used.
from NeuroGraph import preprocess
state_dataset = preprocess.Brain_Connectome_State(root, name, n_rois, threshold, path_to_data, n_jobs)We also offer similar functionalities for constructing dynamic datasets. You can create a dynamic REST dataset from the data stored locally as follows.
from NeuroGraph import preprocess
ngd = Dyn_Prep(fmri, regs, n_rois=100, window_size=50, stride=3, dynamic_length=None)
dataset = ngd.dataset
labels = ngd.labels
print(len(dataset),len(labels))Here the dataset is a list containing dynamic graphs in the form of PyG Batch, which can be easily fed into graph machine learning pipelines. The following examples demonstrate how a dynamic REST dataset can be downloaded and preprocessed on the fly.
from NeuroGraph import preprocess
dyn_obj = preporcess.Dyn_Down_Prep(root, name,s3,n_rois = 100, threshold = 10, window_size = 50,stride == 3, dynamic_length=150)
dataset = dyn_obj.data_dictDyn_Down_Prep class downloads and preprocess the rest dataset and provides a dictionary that contains a list of dynamic graphs against each id. The dataset can be further prprocessed as follows to construct each benchmark.
from NeuroGraph import preprocess
dyn_obj = preporcess.Dyn_Down_Prep(root, name,s3,n_rois = 100, threshold = 10, window_size = 50,stride == 3, dynamic_length=150)
dataset = dyn_obj.data_dict
gender_dataset, labels = [],[]
for k,v in dataset.items():
if v is None:
continue
l = v[0].y
gender = int(l[0].item())
sub = []
for d in v:
new_data = Data(x = d.x, edge_index = d.edge_index, y = gender)
sub.append(new_data)
batch = Batch.from_data_list(sub)
gender_dataset.append(batch)
labels.append(gender)
print("gender dataset created with {} {} number of instances".format(len(gender_dataset), len(labels)))
new_dataset = {'labels':labels, "batches":gender_dataset}
age_dataset, labels = [],[]
for k,v in dataset.items():
if v is None:
continue
l = v[0].y
age = int(l[1].item())
if age <=2: ### Ignoring subjects with age >=36
sub = []
for d in v:
new_data = Data(x = d.x, edge_index = d.edge_index, y = age)
sub.append(new_data)
batch = Batch.from_data_list(sub)
age_dataset.append(batch)
labels.append(gender)
print("Age dataset created with {} {} number of instances".format(len(age_dataset), len(labels)))
new_dataset = {'labels':labels, "batches":age_dataset}
wm_dataset, labels = [],[]
for k,v in dataset.items():
if v is None:
continue
l = v[0].y
wm = int(l[2].item())
if wm is not None: ## there are some None which should be removed
sub = []
for d in v:
# print(d)
new_data = Data(x = d.x, edge_index = d.edge_index, y = wm)
sub.append(new_data)
batch = Batch.from_data_list(sub)
wm_dataset.append(batch)
labels.append(gender)
print("Working memory dataset created with {} {} number of instances".format(len(wm_dataset), len(labels)))
new_dataset = {'labels':labels, "batches":wm_dataset}
fi_dataset, labels = [],[]
for k,v in dataset.items():
if v is None:
continue
l = v[0].y
fi = int(l[3].item())
if not math.isnan(fi): ## there are some None which should be removed
sub = []
for d in v:
# print(d)
new_data = Data(x = d.x, edge_index = d.edge_index, y = fi)
sub.append(new_data)
batch = Batch.from_data_list(sub)
fi_dataset.append(batch)
labels.append(gender)
print("Fluid intelligence dataset created with {} {} number of instances".format(len(fi_dataset), len(labels)))
new_dataset = {'labels':labels, "batches":fi_dataset}