Improved Brain Age Estimation With Slice-Based Set Networks

Abstract

Deep Learning for neuroimaging data is a promising but challenging direction. The high dimensionality of 3D MRI scans makes this endeavor compute and data-intensive. Most conventional 3D neuroimaging methods use 3D-CNN-based architectures with a large number of parameters and require more time and data to train. Recently, 2D-slice-based models have received increasing attention as they have fewer parameters and may require fewer samples to achieve comparable performance. In this paper, we propose a new architecture for BrainAGE prediction. The proposed architecture works by encoding each 2D slice in an MRI with a deep 2D-CNN model. Next, it combines the information from these 2D-slice encodings using set networks or permutation invariant layers. Experiments on the BrainAGE prediction problem, using the UK Biobank dataset, showed that the model with the permutation invariant layers trains faster and provides better predictions compared to other state-of-the-art approaches.

Local Brain-Age: A U-Net Model

“…Out proposed U-Net local brain-age framework has some strengths and weaknesses. Our model was assessed on a large multi-site testing set with a flat distribution of chronological age across the adult lifespan (18–90 years; Supplementary Figure 1 ), a wider interval than a number of studies that rely on UK Biobank (Bintsi et al, 2020 ; Gupta et al, 2021 ; Peng et al, 2021 ) or other narrower-age range studies. Our model showed excellent test-retest reliability, giving confidence that the model could be applied longitudinally to assess individual patterns of brain-ageing changes.…”

“…Some studies have gone further and extracted “patch” level information on brain-age, subsequently averaging predictions across brain regions to arrive at a global-level prediction (Beheshti et al, 2019 ; Pawlowski and Glocker, 2019 ; Bintsi et al, 2020 ; Gupta et al, 2021 ). In Bintsi et al ( 2020 ), the authors use a ResNet (He et al, 2016 ) for each 64 3 3D block, reporting MAE values between 2.16 and 4.19 depending on block origin.…”

“…However, the authors use Support Vector Regression to aggregate the patch-level results to arrive at a global level prediction and do not provide patch-level results in the cortical regions. Similarly, Gupta et al ( 2021 ) proposed a slice-level MRI encoding network, followed by an aggregation method to obtain global-level predictions. Likewise, the authors do not provide results at finer grained scales.…”

Secure Federated Learning for Neuroimaging

“…The difference between the true chronological age and the predicted age of the brain is considered an important biomarker for early detection of age-associated neurodegenerative and neuropsychiatric diseases [33], [34], such as cognitive impairements [35], schizophrenia [36], chronic pain [37]. Recently, deep learning methods based on RNN [38], [39] and CNN [40]- [43] architectures have demonstrated accurate brain age predictions.…”

Secure Neuroimaging Analysis using Federated Learning with Homomorphic Encryption

“…15 Recently, this task has been used for benchmarking, as ground truth (the person's real age) is known. Deep learning methods have been used to predict an individual's brain age both in centralized [16][17][18][19][20] and federated learning settings. 8 In our study, we perform the BrainAge prediction task using a 2D Convolutional Neural Network (CNN), which was shown 20 to yield better predictive performance compared to its 2D-Slice-RNN 18 and 3D-CNN 16, 19 counterparts.…”