DuSCN-FusionNet: An Interpretable Dual-Channel Structural Covariance Fusion Framework for ADHD Classification Using Structural MRI

1. Problem Statement

Attention Deficit Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder, yet its diagnosis lacks reliable, objective neurobiological biomarkers. While Structural MRI (sMRI) offers a non-invasive way to investigate brain alterations, current deep learning (DL) approaches for ADHD classification face two critical limitations:

1. Black-Box Nature: Most DL models lack interpretability, obscuring the anatomical basis of their predictions and limiting clinical trust.
2. Under-utilization of Structural Relationships: Existing methods often focus on raw voxel data or simple regional features, failing to explicitly model inter-regional structural covariance (how different brain regions co-vary in morphology across a population), which is crucial for understanding large-scale brain organization in disorders.

The authors propose a framework that not only achieves high classification accuracy but also provides Explainable AI (XAI) to identify specific brain regions acting as potential biomarkers.

2. Methodology: DuSCN-FusionNet

The proposed framework, DuSCN-FusionNet, is a dual-channel deep learning pipeline designed to process sMRI data through Structural Covariance Networks (SCNs). The architecture consists of four main stages:

A. Data Preprocessing

• Dataset: T1-weighted MRI scans from the ADHD-200 Peking University site (194 subjects: 78 ADHD, 116 Healthy Controls).
• Processing: Images are skull-stripped, spatially normalized, and intensity-normalized using Median and Median Absolute Deviation (MAD) to correct for scanner bias. Values are clipped to $[-3, 3]$ and rescaled to $[0, 1]$.
• Parcellation: The brain is divided into 116 Regions of Interest (ROIs) using the Automated Anatomical Labeling (AAL) atlas.

B. Feature Extraction

For each ROI, two complementary morphometric features are extracted:

1. Mean Normalized Intensity (MNI): Represents the average tissue density/intensity.
2. Inter-Quartile Range (IQR): Captures intra-regional structural heterogeneity (variability).

• Auxiliary Features: Global image statistics (mean, standard deviation, median) are also computed.

C. Structural Covariance Networks (SCNs) Modeling

Instead of feeding raw images to a CNN, the model constructs SCNs to represent inter-regional relationships:

1. Group-Level Covariance: A population-level correlation matrix ($C$) is computed from the feature matrix of all training subjects.
2. Subject-Specific Covariance: An individual subject's feature vector is normalized and converted into an outer-product matrix ($C_{ind}$).
3. Blending: The final subject-specific SCN is a weighted blend of group and individual information: $C(s) = \alpha C + (1-\alpha)C_{ind}(s)$.
4. Dual-Channel Input: Two distinct SCN matrices are generated per subject:
   • Channel 1: Based on MNI (Intensity-based).
   • Channel 2: Based on IQR (Heterogeneity-based).
   • These form a tensor $X(s) \in \mathbb{R}^{2 \times 116 \times 116}$.

D. Classification Module (Late Fusion)

The network employs a late-stage fusion strategy:

1. SCN-CNN Encoder: A Convolutional Neural Network (3 convolutional layers with 64, 128, 256 filters) processes the dual-channel SCN tensor to extract high-level structural covariance features ($z_{SCN}$).
2. Auxiliary Branch: The 119 auxiliary features (ROI-wise IQR + global stats) are processed through a Multi-Layer Perceptron (MLP) to produce $z_{aux}$.
3. Fusion: The vectors $z_{SCN}$ and $z_{aux}$ are concatenated and passed through fully connected layers for final binary classification (ADHD vs. HC).

E. Interpretability (Post-hoc Analysis)

To address the "black-box" issue, the authors adapted Grad-CAM for the SCN domain:

• Gradients of the ADHD class score are back-propagated to the final convolutional layer.
• Since SCNs represent relationships rather than spatial voxels, the activation maps are aggregated across rows and columns to derive ROI-level importance scores.
• ROIs exceeding the 90th percentile are identified as candidate biomarkers.

3. Key Contributions

1. Dual-Channel SCN Representation: The first sMRI-only deep learning pipeline to utilize dual-channel SCNs (intensity-based and heterogeneity-based) to capture both mean morphological changes and intra-regional variability.
2. Late-Stage Fusion Architecture: Integration of global statistical descriptors and ROI variability features alongside the SCN-CNN branch to enhance discrimination.
3. SCN-Adapted Grad-CAM: A novel adaptation of Grad-CAM to the Structural Covariance domain, enabling the localization of specific brain regions (ROIs) that drive the classification decision, thereby providing anatomical interpretability.

4. Experimental Results

The model was evaluated using stratified 10-fold cross-validation with a 5-seed ensemble strategy on the ADHD-200 Peking University site.

• Performance Metrics:
  • Balanced Accuracy: 80.59% (± 9.39)
  • AUC: 0.778
  • Precision: 81.66%
  • Recall: 80.59%
  • F1-Score: 80.27%
• Comparison: The model outperformed several state-of-the-art methods (e.g., 3D MVA-CNN, Dense Attentive GAN) in terms of providing biomarker interpretability, while maintaining competitive accuracy (80.59% vs. 78.8% for Wang et al., 65.86% for Zou et al.).
• Ablation Study: Removing auxiliary features dropped accuracy to ~62%, and removing the seed ensemble dropped it to ~72%, confirming the necessity of both fusion and ensemble strategies.

5. Significance and Biological Insights

• Biomarker Identification: The Grad-CAM analysis identified 12 suprathreshold ROIs as critical for classification. Key regions included:
  • Bilateral Caudate: Critical for executive control and attention.
  • Cingulum (Anterior/Middle/Posterior): Associated with cognitive control and emotion processing.
  • Paracentral Lobules and Cerebellar Vermis (Vermis 9–10): Linked to attentional and affective modulation.
• Clinical Relevance: These findings align with established ADHD neurobiology (fronto-striatal circuitry), validating the model's ability to learn clinically meaningful patterns rather than noise.
• Future Impact: The framework demonstrates that explicitly modeling structural covariance alongside intra-regional heterogeneity yields robust, interpretable models, paving the way for clinically viable AI tools in neuroimaging.