Bayesian Optimization for Neuroimaging Pre-processing in Brain Age Classification and Prediction | JoVE Visualize

Area of Science:

Neuroscience
Machine Learning
Medical Imaging

Background:

Neuroimaging-based age prediction serves as a biomarker for brain aging, cognitive function, health outcomes, and neurodegenerative disease progression.
Current age-prediction algorithms have measurement errors, necessitating improvements in experimental pipelines.
T1-weighted MRI preprocessing, including normalization, resampling, and smoothing, influences age prediction accuracy, with resampling parameters often chosen arbitrarily.

Purpose of the Study:

To enhance brain-age prediction accuracy by optimizing resampling parameters (voxel size and smoothing kernel size) using Bayesian optimization.
To investigate the impact of optimized parameters on both classification of young vs. old brains and regression-based chronological age prediction.
To evaluate the generalizability of the optimized models to an independent dataset.

Main Methods:

Support vector machines were trained on data from 2003 healthy individuals (aged 16-90 years) for classification and regression tasks.
Bayesian optimization was employed to adaptively determine optimal voxel size and smoothing kernel size by evaluating parameter space.
The age-regression model's generalizability was tested on the independent CamCAN dataset (N = 648, aged 18-88 years).

Main Results:

Classification accuracy for distinguishing young (<22 years) from old (>50 years) brains reached 88.1% with optimal parameters (voxel size = 11.5 mm³, smoothing kernel = 2.3 mm).
For chronological age prediction, a Mean Absolute Error (MAE) of 5.08 years was achieved using optimal parameters (voxel size = 3.73 mm³, smoothing kernel = 3.68 mm).
Applying the Bayesian optimization framework to the independent dataset yielded the best performance, outperforming parameters optimized on the initial training set.

Conclusions:

Bayesian optimization can be used to derive case-specific preprocessing parameters for neuroimaging-based brain-age prediction models.
Optimizing preprocessing parameters in specific contexts can improve statistical sensitivity and reduce experimenter bias.
The findings demonstrate the potential for adaptive optimization techniques to enhance the reliability and accuracy of brain aging biomarkers.