Dimensionality reduction in 3D causal deep learning for neuroimage generation: an evaluation study.

Area of Science:

• Medical imaging
• Artificial intelligence
• Neuroscience

Background:

• Causal deep learning (DL) with normalizing flows generates counterfactual images for medical applications like explainability and in-silico studies.
• High-resolution 3D neuroimages require dimensionality reduction (DR) for computationally efficient DL model training.
• The choice of DR method significantly influences the quality and reliability of counterfactual neuroimage generation.

Purpose of the Study:

• To compare the impact of various DR techniques on counterfactual neuroimage generation.
• To identify the optimal DR method for causal DL applications in neuroimaging.

Main Methods:

• Five DR techniques were applied to 23,692 3D brain images: 2D PCA, 2.5D PCA, 3D PCA, autoencoder, and VQ-VAE.
• Causal DL models were trained on the reduced-dimensionality data.
• Convolutional neural networks evaluated age and sex changes in counterfactual images using Mean Absolute Error (MAE) and classification accuracy.

Main Results:

• 2.5D PCA yielded the lowest MAE (4.16) for age alterations.
• Autoencoder embedding achieved the highest sex classification accuracy (97.84%) but increased age MAE to 5.24.
• 3D PCA demonstrated a balanced performance with an age MAE of 4.57 and sex classification accuracy of 94.01% when altering age, and 94.73% accuracy with an age MAE of 3.84 when altering sex.

Conclusions:

• 3D PCA is the most suitable DR method for causal neuroimage analysis.
• The selection of DR technique critically affects the performance of causal DL models in neuroimaging.
• 3D PCA provides a robust balance between accuracy and efficiency for generating reliable counterfactual neuroimages.