Han Peng - CMIC/WEISS joint seminar series

Han Peng,

Nuffield Department of Clinical Neurosciences, University of Oxford

Title:

Accurate brain age prediction with lightweight deep neural networks

Abstract:
Convolution neural network has huge potential for accurate disease prediction with neuroimaging data, but the prediction performance is often limited by training-dataset size and compute memory requirements. To address this, we propose a deep convolutional neural network model, Simple Fully Convolutional Network (SFCN), benchmarked with brain age prediction 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. We compared our 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 99.3% 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).
 
In this talk, I will introduce the rationale behind the simple design and the techniques (aka ‘tricks’) for boosting performance, including data augmentation, pre-training, model regularization, model ensemble and prediction bias correction.
  
Reference:
Accurate brain age prediction with lightweight deep neural networks
Han Peng, Weikang Gong, Christian F. Beckmann, Andrea Vedaldi, Stephen M. Smith
bioRxiv 2019.12.17.879346; doi: https://doi.org/10.1101/2019.12.17.879346
 
Code and pretrained weights
https://github.com/ha-ha-ha-han/UKBiobank_deep_pretrain