Latent Causal Modeling for 3D Brain MRI Counterfactuals

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Limited sample sizes in structural brain MRI studies hinder deep learning model training.
• Generative models can learn MRI data distributions but struggle with out-of-distribution, diverse data generation.
• Causal models for 3D volumes show potential but often yield lower-quality 3D brain MRIs due to high-dimensional challenges.

Purpose of the Study:

• To develop a method for generating high-quality, diverse 3D brain MRI counterfactuals.
• To address limitations of current generative and causal models in neuroimaging.
• To improve the utility of deep learning in structural brain MRI studies.

Main Methods:

• A two-stage approach was proposed, constructing a Structural Causal Model (SCM) in the latent space.
• Stage one utilized a VQ-VAE for compact MRI volume embedding.
• Stage two integrated a causal model into the latent space, employing a three-step counterfactual procedure with a Generalized Linear Model (GLM).

Main Results:

• The proposed method successfully generated high-quality 3D MRI counterfactuals.
• Experiments were conducted on high-resolution MRI data from ADNI and NCANDA datasets.
• The approach demonstrated effectiveness in producing diverse and high-fidelity counterfactual brain MRI data.

Conclusions:

• The novel two-stage latent space causal modeling approach effectively generates high-quality 3D MRI counterfactuals.
• This method overcomes limitations of existing generative and causal models for neuroimaging.
• The findings have implications for enhancing deep learning model training and data augmentation in structural brain MRI research.