2843
Prediction of Chronological Age from Routine T2-weighted Spin-echo Brain Magnetic Resonance Images by Deep Convolutional Neural Network1Seoul National University Hospital, Seoul, Korea, Republic of
Synopsis
Brain-predicted age may be used as a potential biomarker of brain aging. Given that 2D T2-weighted images are more routinely acquired from patients than those 3D images, this study investigated the potential applicability of 2D images in deep learning-based prediction of brain age with an assumption that each individual slice of the T2-weighted brain images possesses brain age-associated features learnable by a convolutional neural network (CNN). The purpose of this study was to investigate whether there are learnable features by a CNN in each slice of routine T2-weighted spin-echo brain MR images that might be associated with normal aging.
INTRODUCTION
Brain-predicted age may be used as a potential biomarker of brain aging. Previous studies reported the feasibility of deep learning-based prediction of brain age, which substantially reduces the computation time for data pre-processing. These previous studies predicted brain age from the whole 3D T1-weighted1 or 3D time-of-flight MR angiography2 images in combination with 3D convolutional filters. There may be age-related features that can be better assessed by T2-weighted images, such as white matter hyperintensities3 or T2 relaxation time changes.4 Given that 2D T2-weighted images are more routinely acquired from patients than those 3D images, this study investigated the potential applicability of 2D images in deep learning-based prediction of brain age with an assumption that each individual slice of the T2-weighted brain images possesses brain age-associated features learnable by a convolutional neural network (CNN). Therefore, the purpose of this study was to investigate whether there are learnable features by a CNN in each slice of routine T2-weighted spin-echo brain MR images that might be associated with normal aging.
METHODS
The study was approved by IRB. T2-weighted spin-echo images of routine brain MRI scan of 500 healthy patients were included from Health Promotion Center of our institution. All MRI scans were performed for routine check-up without evidence of brain disease, and investigator reviewed formal report to exclude scans with abnormal findings. The mean age of dataset was 54.5 years (range; 19-88 years). The MRI scans were performed by using various scanners as follows; GE 3.0T system (n = 231), Siemens 3.0T system (n = 13), Philips 3.0T system (n = 13), GE 1.5T system (n = 200) and Siemens 1.5T system (n = 43). Each T2-weighted series consists of 23 to 31 slices, with 5mm section thickness.
The 500 patients were randomly assigned to a training (n = 400), a validation (n = 50), and a test (n = 50) sets. For each patient those 12 slices that contain mainly brain tissue were included in the data set (Figure 1). Age value of each patient were assigned to every 12 slices as ground-truth. Then, the slices in the training (n = 4800) and the validation (n = 600) sets were randomly shuffled. Images were resampled to 256×256 and zero-center normalized.
A CNN was designed and Bayesian-optimized in Matlab (Mathworks, Inc.).5 For the training of the CNN a stochastic gradient descent with momentum algorithm (SGDM) was used. The loss function was mean-squared-error. The CNN was trained to predict brain age from each slice. For each patient in the test set the mean age over the 12 slices was considered to represent the age of the patient. The performance of the CNN was evaluated by calculating the mean absolute error (MAE) and the Pearson’s correlation coefficient (r) between the ground truth and the CNN-predicted brain ages.
RESULTS
After optimization of hyperparameters of CNN, we conducted training for 100 epochs using batch size of 8. The optimized CNN is shown in Figure 2. From the test set, MAE = 7.11 years and r = 0.86 (p < 0.001) were obtained (Figure 3).DISCUSSION
The MAE and r need to be further improved with more training data and better CNN design and architecture. Nonetheless, our preliminary results may support that each individual slice of the T2-weighted brain images possesses brain age-associated features learnable by a CNN. Upon further improvement, the proposed approach for the prediction of brain age from each slice using 2D convolution may potentially be advantageous over the previously reported approaches from the whole brain using 3D convolution as follows; 1) most of brain MR scans include axial T2-weighted spin-echo images as a routine sequence and readily available, 2) possibility of open black box which section of brain is most important for estimating age, that may feedback to radiologist new insights of routine reading related to brain aging. To achieve this goals, further study is warranted with large size of datasets, and investigation of subsamples, e.g. slice-by-slice comparison analysis of predictive performance. In addition, investigation of clinical significance also should be warranted, including analysis of predicted age difference from chronological age in specific disease group. Finally, external validation including datasets from different institutions should be mandatory to generalize this results.CONCLUSION
Each individual slice of the routine T2-weighted brain MR images might possess brain age-associated features learnable by a CNN.Acknowledgements
No acknowledgement found.References
1. Cole JH, Poudel RPK, Tsagkrasoulis D, et al. Predicting brain age with deep learning from raw imaging data results in a reliable and heritable biomarker. Neuroimage. 2017;163:115-124.
2. Nam Y, Lee J, Kim DH, et al. Predicting the age from time of ight MR angiography using 3D convolutional neural network. Proc. Intl. Soc. Mag. Reson. Med. 26 (2018), 2097.
3. Meyer JS, Kawamura J, Terayama Y. White matter lesions in the elderly. J Neurol Sci 1992;110:1-7.
4. Kumar R, Delshad S, Woo MA, et al. Age-related regional brain T2-relaxation changes in healthy adults. J Magn Reson Imaging. 2012;35(2):300-308.
5. Snoek J, Larochelle H, Adams RP. Practical bayesian optimization of machine learning algorithms. Adv Neural Inf Process Syst. 2012; 2951-2959.
Figures
