Combining anatomical and functional networks for neuropathology identification: A case study on autism spectrum disorder

Sarah Itani, Dorina ThanouMedical Image Analysis 2021 · DOI

TL;DR

This paper introduces a novel approach to identify Autism Spectrum Disorder (ASD) by combining brain structure and function using Graph Signal Processing (GSP). Their method, which extends the Fukunaga-Koontz Transform to graph data, achieves state-of-the-art classification accuracy on the ABIDE dataset, offering a more interpretable diagnostic tool.

What Problem Does This Paper Address?

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition with a growing prevalence, yet its underlying causes and diagnostic markers remain elusive. While previous research has explored brain structure (structural connectivity) and brain activity (functional connectivity) separately, a unified approach that leverages both is crucial for a deeper understanding and more accurate diagnosis. Existing methods often struggle to capture the intricate interplay between these two aspects of brain function, and many advanced machine learning techniques lack interpretability, which is vital for clinical applications. This paper aims to bridge this gap by developing a method that integrates structural and functional brain data for ASD classification, while maintaining interpretability.

Methods

The core of this paper lies in its novel application of Graph Signal Processing (GSP) to neuroimaging data. Here's a breakdown of their approach:

  • Graph Representation: The brain is modeled as a graph where brain regions of interest (ROIs) are nodes, and edges represent the anatomical distance between them. This creates a static structural graph.
  • Graph Signals: Resting-state fMRI (rs-fMRI) signals, representing brain activity, are treated as time-varying "signals" living on the nodes of this structural graph.
  • Graph Fourier Transform (GFT): To analyze the frequency content of these brain signals within the graph structure, the authors employ the GFT. This transforms the time-series data into the graph's spectral domain, revealing how signals propagate across the network at different "graph frequencies."
  • Fukunaga-Koontz Transform (FKT) Extension: A key innovation is the extension of the FKT to this graph-based, time-varying signal data. The FKT is a powerful technique for finding discriminative features between two classes. By adapting it to the graph spectral domain, the authors aim to identify patterns in brain activity that are most distinguishing between individuals with ASD and neurotypical (NT) individuals.
  • Decision Tree Classifier: To ensure interpretability, a decision tree is used to classify subjects based on the discriminative features extracted by the extended FKT. This allows for a clear, rule-based diagnosis.

Datasets

The study utilizes the Autism Brain Imaging Data Exchange (ABIDE I) dataset, a widely used public repository of rs-fMRI data from individuals with and without ASD. Specifically, they used the preprocessed data from the C-PAC pipeline, focusing on the Automated Anatomical Labeling (AAL) atlas with 90 regions of interest (ROIs). The dataset was filtered based on specific inclusion criteria (eyes open during acquisition, age < 18, low framewise displacement) to ensure data quality and consistency, resulting in a subset of 452 subjects (251 NT, 201 ASD).

Key Insights and Ideas

  • Structure-Function Synergy: The paper strongly suggests that combining structural and functional brain information is not just additive but synergistic, leading to significantly better classification performance than using either modality alone.
  • Beyond Simple Frequency Bands: Contrary to some previous work, their findings indicate that ASD-related differences are not confined to specific low, middle, or high graph frequency bands. Instead, discriminative patterns emerge from complex combinations across the entire frequency spectrum.
  • Interpretability is Key: The choice of a decision tree classifier, coupled with the interpretable nature of the extracted graph Fourier features, highlights the importance of explainable AI in clinical applications. This allows clinicians to understand why a diagnosis is made.
  • Topological Influence: The study demonstrates that the choice of brain topology (how nodes are connected) significantly impacts performance, with a carefully chosen "two-nearest neighbor" topology yielding superior results compared to fully connected or random networks.

Key Results

  • State-of-the-Art Performance: The proposed framework significantly outperforms existing methods, including the standard GFT-based approach and the Spatial Filtering Method (SFM), achieving performance gaps of up to 12% and demonstrating statistically significant improvements.
  • Decision Tree Accuracy: Using a decision tree, their method achieves averaged test accuracies close to 75% on the ABIDE dataset, which is highly promising given the complexity of ASD.
  • Frontal and Temporal Lobe Influence: Analysis of the discriminative features reveals the significant involvement of the frontal and temporal lobes in ASD diagnosis, aligning with established neuroscience findings.
  • Adult vs. Adolescent Performance: While performance is generally lower on a smaller adult subsample, their framework still outperforms SFM, suggesting robustness even with limited data.

Limitations and Future Work

  • Simplistic Topology: The authors acknowledge that their graph topology definition (based on anatomical distance and nearest neighbors) is a simplification. Future work could explore more realistic topologies derived from individual Diffusion Tensor Imaging (DTI) data.
  • Sample Heterogeneity: The ABIDE dataset, while large, exhibits heterogeneity in age, multisite origin, and ASD spectrum profiles. The authors suggest that larger, more homogeneous samples would be beneficial for further investigation.
  • Individualized Features: While the current approach identifies general discriminative patterns, future research could explore how these features vary individually, potentially leading to personalized diagnostic markers.
  • Exploring Other Topologies: Further experiments with different graph topologies (e.g., weighted fully connected, uniformly connected) were conducted, showing that the 2-NN topology was most effective, suggesting that non-local interactions might add noise.

Summary updated

BibTeX
@article{itani2021combining,
  author = {Sarah Itani and Dorina Thanou},
  title = {{Combining anatomical and functional networks for neuropathology identification: A case study on autism spectrum disorder}},
  journal = {Medical Image Analysis},
  year = {2021},
  doi = {10.1016/j.media.2021.101986},
}
Figure