CVAE-based ADHD neuroimaging analysis

Project definition

Background

• Reconstruct 3D brain MRIs using CVAEs.
• Disentangle “salient” ADHD-related features from “background” features common to both ADHD and typically developing children.
• Assess how well the learned latent spaces reflect behavioral and clinical variation using:
  • Silhouette Analysis
  • Representational Similarity Analysis (RSA)

Tools

This project used:

• Python (NumPy, Pandas, SciPy, Scikit-learn, Matplotlib, Seaborn)
• Keras (TensorFlow backend) for deep learning
• UMAP for latent space visualization
• Representational Similarity Analysis (RSA) using Kendall’s tau
• Silhouette analysis for latent space separability
• GitHub for version control

Data

This project used the publicly available ADHD-200 dataset, specifically the Burner-preprocessed version:

• Structural MRI data processed via voxel-based morphometry (VBM)
• Normalized 3D gray matter volumes (64×64×64)
• Accompanied by phenotypic variables: age, sex, diagnosis, subtype, medication, IQ, etc.

Deliverables

At the end of the project, we produced:

• A working CVAE framework for modeling neuroanatomical variation
• Visualizations of latent features and synthetic brain reconstructions
• RSA and clustering results relating brain features to clinical data
• Well-documented code and reproducible analysis notebooks

Results

Progress overview

We trained a CVAE model with two latent components:

• s (salient features): ADHD-specific anatomical variation
• z (background features): common/shared variation

We evaluated the model using:

• Silhouette scores for latent clustering
• RSA to correlate latent dimensions with phenotypic variables
• GMM clustering and BIC to test for discrete vs. continuous subtype structure

Tools I learned during this project

• Contrastive representation learning in generative models
• Implementation of CVAEs in Keras
• Preprocessing and working with VBM MRI data
• Representational Similarity Analysis
• Model interpretability techniques for brain data

Results

Deliverable 1: CVAE analysis

• Salient features (s) were significantly correlated with:

  • ADHD Index
  • Inattentive and Hyperactive/Impulsive scores
  • Medication status and age

• Shared features (z) were more related to:

  • IQ, gender, and scan site

Deliverable 2: Clustering analysis

• Gaussian Mixture Models (GMM) with Bayesian Information Criterion (BIC) showed:
  • Lowest BIC at 1 cluster, suggesting continuous heterogeneity
  • Results align with dimensional models of psychiatric disorders

Deliverable 3: Code and notebook

• Full training pipeline in Train-CVAE-ADHD200.ipynb
• Visualization and RSA in helper scripts
• Readme documentation and reproducibility checklist included

Conclusion and acknowledgement

This project demonstrates the potential of contrastive deep generative models like CVAEs to disentangle disorder-specific neuroanatomical features from shared variation. Our findings suggest ADHD may be better described along a spectrum rather than discrete subtypes. We thank the Brainhack School instructors and the open neuroimaging community for providing tools and data that made this project possible.