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

Abstract

Using 3D CNNs on high-resolution medical volumes is very computationally demanding, especially for large datasets like UK Biobank, which aims to scan 100,000 subjects. Here, we demonstrate that using 2D CNNs on a few 2D projections (representing mean and standard deviation across axial, sagittal and coronal slices) of 3D volumes leads to reasonable test accuracy (mean absolute error of about 3.5 years) when predicting age from brain volumes. Using our approach, one training epoch with 20,324 subjects takes 20–50 s using a single GPU, which is two orders of magnitude faster than a small 3D CNN. This speedup is explained by the fact that 3D brain volumes contain a lot of redundant information, which can be efficiently compressed using 2D projections. These results are important for researchers who do not have access to expensive GPU hardware for 3D CNNs.

Brain Age Prediction Using 2D Projections Based on Higher-Order Statistical Moments and Eigenslices from 3D Magnetic Resonance Imaging Volumes

“…In our previous work, we examined the possibility of assessing brain age using deep learning using a limited amount of two-dimensional images derived from brain volume [26], inspired by Langner et al [27], instead of using each full 3D volume. The result was a substantially faster training, about 25 min compared to the typical 48 h or more for using a 3D network.…”

“…The specific images used in our previous work [26] were maps of the mean or standard deviation of values along three axes of the brain volume (transversal, sagittal, coronal). We selected these three projections as they are natural and easy to work with.…”

“…By projection, we here mean a way to obtain a 2D image from a 3D volume. Figure 1 shows an overview of our 2D projection approach, which, compared to our previous method [26], uses more channels per axis. Specifically, we tried adding skewness and kurtosis to the previous mean and standard deviation maps, thus using up to four 2D projections per axis.…”

“…Our dataset consists of 29,035 T1-weighted brain volumes from U.K. Biobank [29][30][31][32], which was also used in our previous work [26]. All subjects were scanned using one of four Siemens Skyra 3T scanners with a Siemens 32-channel RF receive head coil, available in Newcastle upon Tyne, Stockport, Reading and Bristol.…”

“…According to our previous investigations, the standard deviation contained more information than the mean, measured by how much the mean average error differed between otherwise identical models with and without one channel [26]. For this reason, it seemed promising to include higher-order statistical moments.…”