Evaluation of 3D Counterfactual Brain MRI Generation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Counterfactual generation in medical imaging aids disease mechanism understanding and data generation.
• Generating realistic 3D brain MRIs is challenging due to data scarcity and complexity.
• Existing methods lack standardized evaluation protocols for counterfactual brain MRI generation.

Purpose of the Study:

• To adapt six generative models for 3D counterfactual brain MRI generation using an anatomy-guided framework.
• To evaluate these models on their ability to simulate hypothetical changes in brain structure.
• To assess model performance on datasets from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA).

Main Methods:

• Converted six generative models into 3D counterfactual approaches.
• Incorporated an anatomy-guided framework with a causal graph, using regional brain volumes as conditioning inputs.
• Evaluated models on T1-weighted brain MRIs (T1w MRIs) from ADNI and NCANDA datasets.

Main Results:

• Anatomically grounded conditioning effectively modified targeted brain regions.
• Limitations were observed in preserving non-targeted anatomical structures.
• Model generalizability was tested across different datasets.

Conclusions:

• The developed framework enables interpretable and clinically relevant generative modeling of brain MRIs.
• Current models show potential but require further development to capture complex anatomical interdependencies.
• Highlights the need for novel architectures in counterfactual brain MRI generation.