Efficient Brain Age Prediction from 3D MRI Volumes Using 2D Projections.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Neuroscience research

Background:

• 3D Convolutional Neural Networks (CNNs) are computationally intensive for high-resolution medical volumes.
• Large datasets like UK Biobank pose significant hardware challenges for 3D CNN analysis.
• Efficient analysis methods are crucial for large-scale medical research.

Purpose of the Study:

• To develop a computationally efficient method for age prediction from brain volumes.
• To assess the feasibility of using 2D CNNs on projected 3D brain data.
• To enable age prediction analysis on large datasets with limited hardware resources.

Main Methods:

• Utilized 2D CNNs applied to 2D projections (mean and standard deviation) of 3D brain volumes.
• Projections were derived across axial, sagittal, and coronal slices.
• Evaluated performance on a large dataset, comparing speed and accuracy against 3D CNNs.

Main Results:

• Achieved reasonable test accuracy for age prediction (mean absolute error of ~3.5 years).
• Demonstrated a significant speedup: one training epoch with 20,324 subjects took only 20-50 seconds on a single GPU.
• The 2D projection method proved two orders of magnitude faster than a small 3D CNN.

Conclusions:

• 2D CNNs on projected 3D brain volumes offer a computationally efficient alternative to 3D CNNs.
• This approach democratizes brain volume analysis for researchers lacking expensive GPU hardware.
• The findings facilitate large-scale neuroimaging studies, such as those within UK Biobank.