Equitable modelling of brain imaging by counterfactual augmentation with morphologically constrained 3D deep...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Generative models are crucial for data augmentation in medical imaging.
• Existing methods struggle with data imbalance, distributional shifts, and demographic bias.
• Improving fairness and fidelity in brain image analysis is a critical challenge.

Purpose of the Study:

• To introduce CounterSynth, a conditional generative model for creating biologically plausible, label-driven deformations in brain images.
• To synthesize counterfactual data augmentations addressing limitations in downstream discriminative tasks.
• To enhance model performance and equity across diverse subpopulations.

Main Methods:

• Developed a conditional generative model for diffeomorphic deformations.
• Utilized voxel-based morphometry, classification, and regression to evaluate synthesized counterfactuals.
• Employed Fréchet Inception Distance (FID) for quality assessment.
• Benchmarked CounterSynth on UK Biobank and OASIS MRI datasets.

Main Results:

• CounterSynth successfully synthesizes biologically plausible, label-driven changes in volumetric brain images.
• Demonstrated state-of-the-art improvements in both fidelity and equity of downstream discriminative models.
• Effectively addressed challenges posed by engineered demographic imbalance and confounding.

Conclusions:

• CounterSynth offers a powerful solution for generating high-fidelity, equitable counterfactual data augmentations.
• The model significantly enhances the robustness and fairness of AI in neuroimaging.
• Code availability facilitates further research and application in medical AI.