Brain age prediction using multi-hop graph attention module(MGA) with convolutional neural network

Heejoo Lim, Yoonji Joo, Eunji Ha, Yumi Song, Sujung Yoon, In Kyoon Lyoo, Taehoon Shin

Keywords: Brain age prediction, graph attention, self attention, deep learning

TL;DR: We propose a novel multi-hop graph attention module (MGA) for brain age prediction which allows CNN to learn direct and indirect connections of feature maps.

Abstract: We propose a multi-hop graph attention module (MGA) that addresses the limitation of CNN in capturing non-local connections of features for predicting brain age. MGA converts feature maps to graphs, calculates distance-based scores, and uses Markov property and graph attention to capture direct and indirect connectivity. Combining MGA with sSE-ResNet18, we achieved a mean absolute error (MAE) of 2.822 years and Pearson's correlation coefficient (PCC) of 0.968 using 2,788 T1-weighted MR images of healthy subjects. Our results present a possibility of MGA as a new algorithm for brain age prediction.