DFC_ADNIDataset

​

Overview

​

Class Definition

class DFC_ADNIDataset(InMemoryDataset)

​

Constructor

DFC_ADNIDataset(
    root,
    threshold=0.2,
    label_csv='TADPOLE_Simplified.csv',
    var_name='dynamic_fc',
    transform=None,
    pre_transform=None
)

​

Parameters

​

Attributes

​

Methods

​

fc_to_graph

fc_to_graph(
    matrix,
    node_features=None
) -> Data

• matrix (numpy.ndarray): FC matrix of shape (N, N) for a single time point
• node_features (numpy.ndarray, optional): Node feature matrix. If None, uses identity matrix

• Data: PyTorch Geometric Data object with x, edge_index, and edge_attr

• Uses upper triangle without self-loops to avoid duplicate edges
• Creates undirected graph by adding both (i, j) and (j, i) edges
• Edge attributes represent correlation strengths

​

get

get(idx) -> Data

• idx (int): Index of the graph to retrieve

• Data: Graph data object with subj_id attribute attached

​

load_subject_labels

load_subject_labels(
    label_csv_path,
    label_col='Label_CS_Num'
) -> dict

• label_csv_path (str): Path to CSV file with subject labels
• label_col (str): Column name containing the label values

• dict: Dictionary mapping subject IDs to their labels

​

Data Format

• x: Node feature matrix (default: identity matrix)
• edge_index: Graph connectivity in COO format (undirected)
• edge_attr: Edge weights (correlation values)
• y: Subject label (diagnosis category)
• time_index: Time window index within the scan
• subj_id: Subject identifier (format: sub-{SUBJECT_ID}_run-{RUN_NUM})

​

Usage Example

import torch
from DFC_ADNIDataset import DFC_ADNIDataset
from torch_geometric.loader import DataLoader

# Initialize dataset
dataset = DFC_ADNIDataset(
    root='data/adni',
    threshold=0.25,
    label_csv='TADPOLE_Simplified.csv',
    var_name='dynamic_fc'
)

print(f"Total number of graphs (all time points): {len(dataset)}")
print(f"Number of features: {dataset.num_features}")
print(f"Number of classes: {dataset.num_classes}")

# Access a single time-point graph
graph = dataset[0]
print(f"Subject ID: {graph.subj_id}")
print(f"Time index: {graph.time_index.item()}")
print(f"Label: {graph.y.item()}")
print(f"Nodes: {graph.num_nodes}, Edges: {graph.num_edges}")

# Group graphs by subject
from collections import defaultdict

subject_graphs = defaultdict(list)
for i in range(len(dataset)):
    graph = dataset[i]
    subject_graphs[graph.subj_id].append(i)

print(f"\nNumber of unique subjects: {len(subject_graphs)}")

# Example: Get all time points for a specific subject
for subj_id, indices in list(subject_graphs.items())[:1]:
    print(f"\nSubject {subj_id} has {len(indices)} time points")
    for idx in indices:
        g = dataset[idx]
        print(f"  Time {g.time_index.item()}: {g.num_edges} edges")

# Create a DataLoader
loader = DataLoader(dataset, batch_size=64, shuffle=True)

for batch in loader:
    # Process batch of time-point graphs
    print(f"\nBatch contains:")
    print(f"  {batch.num_graphs} graphs")
    print(f"  {batch.num_nodes} total nodes")
    print(f"  {batch.num_edges} total edges")
    break

​

File Format Requirements

​

Dynamic FC Matrix Files

• A numpy array with key matching var_name parameter (default: 'dynamic_fc')
• Array shape: (T, N, N) where:
  • T = number of time windows
  • N = number of brain regions
• Values represent correlation coefficients between regions at each time window

​

Label CSV Format

• Subject: Subject identifier (underscores removed for matching)
• Label_CS_Num: Numeric label (diagnosis category)

Subject,Label_CS_Num
002S0413,0
002S0559,1
003S1059,2

​

Notes

• Each time window from a DFC matrix becomes a separate graph in the dataset
• If a subject has T time windows, they contribute T graphs to the dataset
• DFC matrices are automatically symmetrized at each time point: (matrix + matrix.T) / 2
• Edges below the threshold are filtered out to create sparse graphs
• The dataset uses upper triangle indexing to avoid duplicate edges in undirected graphs
• Subject IDs are preserved in subj_id_list for efficient subject-based indexing
• When no label is found for a subject, it defaults to 0
• Processed data is saved to data_dfc.pt in the processed/ directory

​

Backward Compatibility