Accurate brain age prediction with lightweight deep neural networks.

Peng H, Gong W, Beckmann CF, Vedaldi A, Smith SM. · 2021

Read Accurate brain age prediction with lightweight deep neural networks. by Peng H, Gong W, Beckmann CF, Vedaldi A, Smith SM. (2021) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc7610710.txt.

Introduction

Peng, Gong, Beckmann, Vedaldi and Smith present a light deep network for predicting brain age from MRI. Deep learning can predict disease from brain scans. Small datasets and memory limits often hold it back.

Their model is called Simple Fully Convolutional Network. It has fewer parameters than popular deep models. That suits small samples and 3D volume data. Tricks included data augmentation and pre-training steps. Ensembles and bias correction also boosted results. On UK Biobank data with 14,503 people, error was 2.14 years. Sex classification reached 99.5% accuracy there. The model won a 2019 prediction challenge as well. Code was released for other imaging tasks.

Catalog record

CREATORPeng H, Gong W, Beckmann CF, Vedaldi A, Smith SM.
YEAR2021
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC7610710

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Accurate brain age prediction with lightweight deep neural networks By Han Peng, Weikang Gong, Christian F. Beckmann, Andrea Vedaldi, Stephen M. Smith Published in: Medical image analysis DOI: 10.1016/j.media.2020.101871 ABSTRACT Deep learning has huge potential for accurate disease prediction with neuroimaging data, but the prediction performance is often limited by training-dataset size and computing memory requirements. To address this, we propose a deep convolutional neural network model, Simple Fully Convolutional Network (SFCN), for accurate prediction of brain age using T1-weighted structural MRI data. Compared with other popular deep network architectures, SFCN has fewer parameters, so is more compatible with small dataset size and 3D volume data. The network architecture was combined with several techniques for boosting performance, including data augmentation, pre-training, model regularization, model ensemble and prediction bias correction. We compared our overall SFCN approach with several widely-used machine learning models. It achieved state-of-the-art performance in UK Biobank data (N = 14,503), with mean absolute error (MAE) = 2.14y in brain age prediction and 99.5% in sex classification. SFCN also won (both parts of) the 2019 Predictive Analysis Challenge for brain age prediction, involving 79 competing teams (N = 2,638, MAE = 2.90y). We describe here the details of our approach, and its optimisation and validation. Our approach can easily be generalised to other tasks using different image modalities, and is released on GitHub. INTRODUCTION The emergence of machine learning techniques has made automatic disease prediction from medical imaging data possible. The recent development of deep learning pushes prediction accuracy beyond human performance in some scenarios, and is able to assist clinical diagnosis/treatment decisions (De Fauw et al., 2018; Kohl et al., 2018; LeCun et al., 2015). In neuroimaging, deep learning has had successes in several applications in predictive and diagnostic analysis, such as brain age prediction and modelling (Cole et al., 2017; Kawahara et al., 2017), sex classification (Arslan et al., 2018), disease prediction (Baumgartner et al., 2018; Korolev et al., 2017; Liu et al., 2018) and brain lesion segmentation (Kamnitsas et al., 2017), yet still faces several challenges. For example, 3D neuroimaging data requires much more GPU memory than most 2D images, meaning that models successful in 2D data (e.g., ImageNet classification (Krizhevsky et al., 2012; Simonyan and Zisserman, 2014)) are infeasible in the 3D scenario. There are several researches mitigating this issue (e.g.) by downsampling the input (Korolev et al., 2017), taking patches (Kamnitsas et al., 2017; Liu et al., 2018) or 2D slices (Bashyam et al., 2020; Lin et al., 2018) as input instead of the 3D full brain, or using a reversible architecture (Brügger et al., 2019), yet involving trade-offs between the GPU memory restriction and the information/performance loss. Further, deep networks usually require a large sample size for model fitting, but neuroimaging datasets often have relatively few samples compared to existing million-sample natural image datasets (Raghu et al., 2019; Russakovsky et al., 2015), which could limit the ability to learn image features effectively, and result in overfitting problems. New model architecture design is needed to address these challenges for neuroimaging applications. Predicting chronological age based on structural brain magnetic resonance imaging (MRI) data shares common challenges with many other neuroimaging applications, and can be used to develop and test deep learning algorithms. It also receives attention for its potential clinical and biological relevance (Ashburner, 2007; Brown et al., 2012; Cole et al., 2018; Cole and Franke, 2017; Davatzikos et al., 2009; Franke et al., 2014, 2010; Franke and Gaser, 2019; Habes et al., 2016; Kaufmann et al., 2019; Neeb et al., 2006). The predicted age can be considered to be the "brain age”, because it is derived purely from the brain imaging data. After estimating brain age, a further quantity of interest is the difference between the predicted age (brain age) and the actual age, sometimes referred to as the brain-age delta. Positive delta implies that a subject’s brain looks older than their actual age, i.e., they are experiencing accelerated ageing. For example, existing studies have observed that the brain-age delta is an effective biomarker that shows differences between different clinical groups (Kaufmann et al., 2019) and is predictive for mortality (Cole et al., 2018). Achieving accurate brain age prediction is an essential pre-requisite for optimising brain-age delta as a biomarker. To reach this goal, many studies have used different models, such as regularized linear regression, support vector machines and Gaussian process regression, for brain age prediction (Franke and Gaser, 2019). Some studies have used deep learning methods (Cole et al., 2017; Feng et al., 2019; Kolbeinsson et al., 2019). However, challenges exist for further improvement of prediction accuracy, especially on small datasets, and some studies have shown that deep learning performs no better than simpler machine learning models in typical neuroimaging datasets (He et al., 2020; Schulz et al., 2019). It has not yet, for example, been clearly established whether more complex deep learning models perform better than simpler models (for the task of brain age prediction using structural MRI data). In addition, predicted brain age is often systematically biased towards the group mean value, resulting in a correlation between the delta and the chronological age, which weakens the validity of the delta as a biomarker (Smith et al., 2019). Therefore, it is both methodologically interesting and scientifically important to develop unbiased high-performance deep learning strategies for brain age prediction. In this paper, a lightweight deep learning architecture, Simple Fully Convolutional Network (SFCN), is presented for brain age prediction. Its architecture is based on the fully convolutional network (FCN) (Long et al., 2015) and the VGG net (Simonyan and Zisser-man, 2014) and takes 3D minimally-preprocessed T1 brain images (and/or preprocessed segmentation outputs) as input. The successful CNN architecture VGG net and its variant with BatchNorm layers (Ioffe and Szegedy, 2015) provide a deep architecture consisting of a sequence of basic blocks: (Conv-BatchNorm-Activation)xN-Pooling. To reduce memory requirements, SFCN keeps only one conv-layer before each MaxPool layer. In addition, we remove all the fully-connected layers, which not only greatly reduces the number of parameters, but also provides a fully convolutional architecture that is versatile for accommodating different input sizes (Long et al., 2015). Using proper data augmentation and regularization techniques, the model achieved state-of-the-art mean absolute error (MAE) of 2.14 years in the UK Biobank dataset (14,503 subjects, of which 12,949 are used for training). This model performs better than several widely-used machine learning models in the literature. In addition, we propose a model ensemble strategy that averages the outputs of deep learning models based on different kinds of preprocessing applied to the T1 data, namely, white matter segmentation, grey matter segmentation, linearly-registered raw T1 and nonlinearly registered raw T1; this further boosts the accuracy of brain age prediction. Finally, we extended the bias correction techniques proposed by (Smith et al., 2019) to greatly reduce the correlation between the brain-age delta and the chronological age, with very little compromise of performance, even when the true ages of test (validation) subjects are unknown. The ensemble SFCN came first in the Predictive Analysis Challenge 2019 in brain age prediction (MAE = 2.90 years) among the 79 participating teams1 With bias correction, our model achieved an MAE of 2.95 years, thereby ranking first also in the other part

Read the complete text