Going deeper with brain morphometry using neural networks

Brain morphometry from magnetic resonance imaging (MRI) is commonly used for estimating imaging biomarkers for many neurodegenerative diseases, including Alzheimer's. Recent work showed that deep convolutional neural networks could estimate morphometric measurements directly from 3D brain MRI within a few seconds, but with limited accuracy, especially for mean curvature and thickness. In this paper, we propose a more accurate and efficient neural network model for brain morphometry named HerstonNet: we developed a 3D ResNet-based neural network to learn rich features directly from MRI, designed a multi-scale regression scheme by predicting morphometric measures at different resolutions, and applied a robust optimization method to avoid poor quality minima, resulting in lower prediction error variance. HerstonNet outperforms the existing approach by 24.30% in terms of intraclass correlation coefficient (agreement measure) to FreeSurfer silver-standard while maintaining a competitive run-time.