Brain-age prediction

Xu, Yue

Permalink

https://hdl.handle.net/2142/104048 Copy

Description

Title

Brain-age prediction

Author(s)

Xu, Yue

Contributor(s)

• Varatharajah, Yogatheesan
• Kalbarczyk, Zbigniew T.

Issue Date

2019-05

Keyword(s)

• Brain Age
• Convolutional Neural Network
• Machine Learning
• U-net
• Image Processing
• MRI images

Abstract

Brain age is a popular measure used in the study of brain aging that estimates the biological age of a brain based on the extent of cerebral atrophy. Accurate brain-age prediction models have wide applicability in the clinical domain, e.g., predicting age-related neurodegenerative diseases. Brain-age models are usually developed by learning the relationship between chronological age and brain structure in samples of healthy individuals with the reasoning that a healthy individual's brain age will be close to his or her chronological age. Furthermore, brain-structure-related features are derived based on structural magnetic resonance imaging (SMRI) data registered to standard atlases and extracting volumetric measures of standard brain regions. The primary goal of this project was to evaluate the efficacy of a brain-age-prediction model that used raw SMRI images instead of preprocessed features. A model that uses raw images to make brain-age predictions can eliminate the need for domain expertise in the development of predictive models, and also identify new biomarkers related to certain neurodegenerative diseases. Therefore, we developed a convolutional-neural-network-based model for predicting brain age based on raw SMRI images. We first processed the images so that all images are of the same scale, size, and orientation. Then we implemented a U-net feature extractor to automatically learn features from processed images. The features learned by the U-net model were used as input to a fully connected neural network to make brain-age predictions. We used the data of healthy individuals (input: raw SMRI images; output: brain age; ground truth: chronological age) to train the network and evaluated the model performance using the mean absolute error (MAE) between predicted brain ages and true ages in a test dataset containing more healthy individuals.

Type of Resource

text

Language

en

Permalink

http://hdl.handle.net/2142/104048

Owning Collections