ICAM-reg

ICAM-reg performs interpretable phenotype mapping of brain imaging by using a VAE-GAN ICAM model to enable simultaneous classification, regression, and feature attribution for neurological phenotype analysis.

Key Features:

• Generative Deep Learning (VAE-GAN / ICAM): Uses a Variational Autoencoder-Generative Adversarial Network (VAE-GAN) implementation of the ICAM model to translate brain imaging data and disentangle class-relevant features from background variation.
• Simultaneous Classification and Regression: Performs classification and regression tasks concurrently to predict diagnostic labels and continuous phenotype measures from individual scans.
• Feature Attribution (FA) Maps: Produces feature attribution (FA) maps that indicate image regions contributing to model predictions and help explain outlier predictions.
• Latent Space Disentanglement via Regression Module: Incorporates a regression module to improve disentanglement of latent representations and isolate phenotype-relevant factors from confounds.

Scientific Applications:

• Cognitive Test Score Prediction: Applied to predict Mini-Mental State Examination (MMSE) scores in the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort.
• Brain Age Prediction: Applied to brain age prediction across neurodevelopmental and neurodegenerative contexts using datasets including the developing Human Connectome Project (dHCP) and UK Biobank.

Methodology:

Train a VAE-GAN ICAM model for image translation to separate disease-relevant features from background variations, include a regression module to enhance latent-space disentanglement, and generate feature attribution (FA) maps to localize predictive contributions.