Leveraging medical images and deep learning to characterize biological age
Approved Research ID: 52887
Approval date: September 13th 2019
We aim to predict the age of participants by using deep learning and medical images. This will allow us to identify if people who look older than they actually are are at a higher risk for the onset of different diseases. Estimating the biological age of an individual can help predict the onset of age-related diseases or determine "biological age". It is possible to approximate the biological age by training a model to learn the chronological age of a large cohort of individuals. The predicted age based on the model is known as biological age. The difference between the chronological and biological age is known as age acceleration. In this project, we aim to predict the biological age of organ systems (e.g., heart, pancreas, bone, liver, and brain), from their image data. Specifically, we will first split the image samples for each tissue into a discovery, test, and validation cohort. Second, we will build a deep learning model (convolutional neural network) to predict and validate chronological age of UK Biobank participants based on the image information from these diverse tissues and we will evaluate the prediction accurate based on the test datasets. We will do this iteratively for each tissue. Third, we will estimate the co-variance of biological age between tissues - for example, is heart tissue age correlated with pancreas tissue age. Finally, we will use the new biological age as a phenotype to execute a novel genome-wide association study to determine genetic variants that are associated with our new phenotype of biological age.
1.) Our first aim is to leverage deep learning and transfer learning to predict the age of the participants using heart, liver, pancreas, brain, and bone image data.
2.) Our second aim is to identify what anatomical features in the image allow the model to predict the age of the participants.
3.) Our third aim is to investigate which features are associated with accelerated agers (participants whose predicted age is lower than their actual age) and decelerated agers (participants whose predicted age is higher than their actual age).
4.) Our fourth aim is to perform a GWAS to identify the SNPs associated with either decelerated or accelerated aging.
1.) We add to aim 1 to predict the age of participants using physical activity data.