Premium
Comparison of deep learning methods for brain age prediction
Author(s) -
Lam Pradeep,
Zhu Alyssa,
Salminen Lauren,
Thomopoulos Sophia I,
Bright Joanna,
Jahanshad Neda,
Thompson Paul M
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.046763
Subject(s) - biobank , test set , artificial intelligence , medicine , cross validation , test (biology) , convolutional neural network , statistics , computer science , mathematics , bioinformatics , paleontology , biology
Background BrainAge ‐ an estimate of a person’s age predicted from their structural MRI ‐ is a biomarker used to quantify age‐related neurodegeneration. The deviation of BrainAge from chronological age (Predicted Age Difference ‐ PAD) is of interest for prognostic research. Here we evaluated the accuracy of several deep learning methods (convolutional neural networks; CNNs) in predicting brain age. Method All experiments were conducted with a subset of 10,801 neurologically healthy participants from the UK Biobank. Of these, 9,301 were scanned at a single site; an additional 1,500 were from auxiliary sites with the same calibration standards. Participants’ age range was 48.9‐76.6 years. The primary site data had a mean age of 62.4 (SD, 7.5; 47% women/53% men). The auxiliary site scanned subjects with a mean age of 63.9 (SD, 7.3; 55% women/45% men). From the 9,301 single‐site scans, 7,153 were used to train the models; the rest were split into validation and testing sets. All auxiliary site data was used for out‐of‐site testing. Each model was trained using a different divergence penalty. In addition, a single model was evaluated on 25,756 participants from the UK Biobank dataset. K ‐fold Cross‐Validation was used to prevent overlap between train, validation, and test sets. For each training instance, 12,365 subjects were used for the training set; the rest were split evenly between validation and test sets. For this experiment, subjects were not filtered out based on neurological health. Result All models performed well on both the in‐site and out‐of‐site test sets with an average mean absolute error (MAE) of 2.98 and 2.94 years, respectively. The model evaluated on 25,756 subjects received a MAE of 2.85 (R 0.845) with a residual correlation of 0.432 between PAD and true age. Conclusion Good accuracy achieved by all models suggests the value of BrainAge as a holistic biomarker of brain aging.