More Robust MRI Preprocessing Yields More Accurate and Generalizable Age Predictions

1467 

Abstract Submission 

Reza Rajabli¹, Mahdie Soltaninejad¹, Sofia Fernandez Lozano¹, Neda Shafiee¹, Vladimir Fonov¹, Danilo Bzdok¹, D Louis Collins¹

¹McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University, Montreal, Quebec

Reza Rajabli  
McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University
Montreal, Quebec

Mahdie Soltaninejad  
McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University
Montreal, Quebec

Sofia Fernandez Lozano  
McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University
Montreal, Quebec

Neda Shafiee  
McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University
Montreal, Quebec

Vladimir Fonov  
McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University
Montreal, Quebec

Danilo Bzdok  
McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University
Montreal, Quebec

D Louis Collins  
McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University
Montreal, Quebec

Deep learning approaches have revolutionized the way various patterns are found in different data settings. The vast number of trainable parameters in deep models make them capable of capturing complex patterns. However, this comes at the expense of their considerable appetite for data. Since these are high variance models, they risk being overtrained if the sample size is not sufficient, which is often the case in medical datasets. For example, training a model on Magnetic Resonance Images (MRIs) to predict a diagnostic/prognostic label, the number of available images can be at most in the order of 100K, significantly less than the 14M samples of ImageNet. Consequently, it is vital to implement measures that ensure the generalizability of medical image deep learning models. A noteworthy approach is to reduce the number of trainable parameters with shallower models. The Simple Fully Convolutional Network (SFCN) [Peng 2021] is one such successful example incorporated into the SFCN-Reg architecture [Leonardsen 2022] to build a generalizable brain age model. Leonardsen et al. trained their model on ~53K MRIs (from multiple studies) reporting a mean absolute error (MAE) of 2.47 years on the internal test and 3.90 years on average on external test images from other studies. While they are one of few who tested their model externally and reported one of the lowest prediction errors, the disparity between internal and external test results raises the question of the model's ability to generalize to future brain scans and maintain high accuracy. We hypothesized that enhanced MR preprocessing could address this issue and lead to a reduction in external test error.

We downloaded 39,676 images from the UK Biobank (UKBB) dataset (version 49190) and applied the following preprocessing steps to each image: 1) brain extraction using SynthStrip [Hoopes 2022], 2) intensity normalization through histogram matching [Shah 2011], 3) denoising with adaptive non-linear means [Manjón 2010], 4) N4 bias field correction [Tustison 2010], 5) repeating step 2, and 6) affine registration to the MNI152 nonlinear symmetric template using ANTs [Avants 2014]. Next, we manually inspected all preprocessed images, and eliminated 1,975 (~5%) as failed, mostly due to either corrupted brain masks or failed registration. Then, we re-implemented the SFCN-Reg architecture with identical hyperparameters as detailed in [Leonardsen 2022], training it on 33,724 randomly subsampled MRIs from the preprocessed images. Finally, we assessed the trained model internally using ~1K MRIs not seen during training from the UKBB and externally (out of domain) on baseline visit scans of cognitively healthy subjects in the ADNI, AIBL, and OASIS3 datasets that underwent the same preprocessing steps and quality control procedures as the UKBB MRIs.

After training our model, we identified the epoch with the best validation error and proceeded to assess its generalizability. To have a better understanding of the model's reliability and robustness, we calculated the mean and standard deviation for the MAE using bootstrapping by randomly selecting 100 subjects 10 times from each test set.

The MAE for predictions on the internal test set (UKBB) was 2.22±0.17 years, and for external tests (ADNI, AIBL, and OASIS3), the MAE results were 3.45±0.28, 3.64±0.26 and 2.82±0.18, respectively (Figs 1-A, 1-B, 1-C, and 1-D).

Since SFCN-Reg was originally trained and tested on approximately ~53K MRIs (3-95 years) from multiple datasets, a direct comparison poses challenges. Nevertheless, we trained the SFCN-Reg on a smaller dataset with narrower age range (45-82 years) while maintaining the hyperparameters consistent with the original paper (Table 1). The experiments demonstrate that a more robust preprocessing pipeline, compared to simpler alternatives used by Leonardson, enhances the generalizability of shallower models.

Aging

Classification and Predictive Modeling ¹

Methods Development

Motion Correction and Preprocessing ²

Other - Brain Age Prediction, Structural MRI, Simple Fully Convolutional Network, Preprocessing