Accurate brain age prediction with lightweight deep neural networks
Han Peng1, Weikang Gong2, Christian F Beckmann3
1Wellcome Centre for Integrative Neuroimaging (WIN FMRIB), University of Oxford, Oxford, OX3 9DU, United Kingdom; Visual Geometry Group (VGG), University of Oxford, Oxford, OX2 6NN, United Kingdom; Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen, Nijmegen, 6525 EN, the Netherlands.
Medical Image Analysis
|November 16, 2020
Summary
We developed a Simple Fully Convolutional Network (SFCN) for accurate brain age prediction using MRI scans. SFCN achieves state-of-the-art results, even with limited data, outperforming other models.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Medical Diagnostics
Background:
- Deep learning shows promise for disease prediction using neuroimaging data.
- Performance is often hindered by limited training data and computational demands.
Purpose of the Study:
- To introduce the Simple Fully Convolutional Network (SFCN), a deep learning model for precise brain age prediction.
- To demonstrate SFCN's effectiveness with T1-weighted structural MRI data.
- To address limitations of dataset size and memory requirements in current models.
Main Methods:
- Proposed SFCN, a deep convolutional neural network with fewer parameters for better compatibility with small datasets and 3D data.
- Employed techniques like data augmentation, pre-training, regularization, ensembling, and bias correction.
- Compared SFCN against established machine learning models.
Main Results:
- Achieved state-of-the-art brain age prediction (MAE = 2.14 years) and sex classification (99.5%) on UK Biobank data (N=14,503).
- Won the 2019 Predictive Analysis Challenge for brain age prediction (N=2,638, MAE = 2.90 years).
Conclusions:
- SFCN offers a robust and efficient approach for brain age prediction from MRI.
- The model's architecture and optimization techniques are generalizable to other neuroimaging tasks.
- The SFCN approach is publicly available on GitHub.
