Synopsis
Heterogeneity
of diffusion MRI data limits the diffusion-MRI-based machine learning model to be
generalized to the data acquired at other sites. To generalize the brain age
model based on diffusion-MRI-derived features, we used transfer learning
techniques to transfer the pre-trained model from the source domain to the
target domain with a few tuning data. We found that 75 tuning data with
transfer learning framework achieved the acceptable performance, and 150 tuning
data achieved the performance comparable to the maximum samples in the target
domain. This study provides a practical solution to solve the limitation of diffusion-MRI-based
model using transfer learning.
Introduction
Brain age is
an emerging imaging biomarker that could be predicted on individuals based on
neuroimaging data using machine learning approaches to model trajectories of
brain aging[1]. Diffusion MRI (dMRI) is suitable for the investigation
of cortical connection and can provide diffusion indices that reflect
structural integrity within interconnected networks[2,3]. Using dMRI
datasets, dMRI-based brain age is a useful aging biomarker to represent the aging
state of white matter[4]. However, the dMRI suffers from between-scanner
variations that hinder direct comparisons across different imaging sites[5].
The dMRI-based brain age model is, therefore, limited within the source domain
and hard to generalize to the data acquired from other sites. To solve the
limitation of dMRI, we aimed to generalize the dMRI-based brain age model using
transfer learning (TL) techniques. Specifically, TL applies a brain age model already
trained in the source domain to a new target domain by tuning the model
parameters with a few data from the target domain.Methods
A large
cohort as the source domain data used to pre-train a brain age predictive model
were obtained from the CamCAN repository[6,7] that consisted of 615
healthy subjects whose age ranged 18-88 years. These data sets, including
T1-weighted images and two-shell diffusion tensor images, were split into the
training (n=500, called CC-training) and testing (n=115, called CC-testing)
sets to develop and evaluate the brain age model, respectively. The data in the
target domain were collected from in-house private MRI database that included
300 tuning data at maximum (NTU-tuning) for TL and 100 testing data
(NTU-testing) for evaluating the TL performance. The data from the target
domain consisted of T1-weighted imaging and diffusion spectrum imaging
datasets. To obtain white matter features to predict brain age, the regularized
version of diffusion MAP-MRI framework was used to reconstruct diffusion image
data into 5 diffusion indices, such as fractional anisotropy[8]. Whole
brain tract-specific analysis was conducted to sample the features according to
the predefined 76 tracts from each diffusion index[9]. These
tract-specific features from the CC-training data were used to create a brain
age pre-trained model by cascade neural network (Fig.1).
To decide the
strategy most effective to adjust the model and the optimal sample size of
tuning data, we proposed three transfer strategies to adjust the model so the
model predicted brain age in the target domain successfully. The strategies
included “co-train” (pool the NTU-tuning and the CC-training data together to
train the model), “TL” (re-train the model with the NTU-tuning data and using
the parameters of the pre-trained model as the model initial values, no data
from the source domain was used) and “TL with co-train” (combine the above two
methods). The “single” framework (train the model using the single NTU-tuning
data, without TL and/or co-training) was performed as the baseline. The
simulated sample size of the tuning data comprised of 30,60,90,120,150,240 and
300 (maximum). The permutation of strategy and sample size would simulate
repeatedly for 30 times to estimate the optimal combination. The simulated result
was evaluated by applying the transferred model to the NTU-testing data through
the statistical metrics including Pearson correlation coefficient and mean
absolute error (MAE). After deciding the optimal combination of tuning sample
size and strategy, we further improved the performance of transferred model by
fine-tuning the hyperparameters and using the advanced optimizer.Results
The pre-trained model predicted individuals’
age for both the CC-training and CC-testing sets with satisfactory performance
(Fig.2). In the simulation, TL strategy achieved better performance than did
the other methods (Fig.3). The performance of transferred model with 150 samples
of NTU-tuning data (correlation=0.922, MAE=5.56 years) was comparable to the
model trained by maximum 300 NTU-tuning data (correlation=0.931, MAE=5.29
years). If we considered the MAE which was one year more than the standard MAE
as the acceptable performance, 75 NTU-tuning data with TL strategy was an
acceptable combination (correlation=0.894, MAE=6.26 years) to transfer the
pre-trained model. After fine-tuning the transferred model, it attained satisfactory
performance as assessed by the NTU-testing data (Correlation=0.916, MAE=5.55
years)(Fig.4).Discussion and Conclusion
We generalized the dMRI-based brain age predictive model to the new
data domain using TL. The TL strategy was an effective paradigm to transfer the
pre-trained model. We found that 75 image data as the tuning set achieved the
acceptable performance compared to the model trained with the 300 datasets in
the target domain (3/4 data-saving). Also, the performance of transferred model
using 150 tuning data was comparable to the standard performance (1/2 data-saving).
This study provides a transfer learning solution to address the problem of
generalization in dMRI-based brain age model.Acknowledgements
The CamCAN data collection and sharing for this study was provided by
the Cambridge Centre for Ageing and Neuroscience (CamCAN). CamCAN funding was
provided by the UK Biotechnology and Biological Sciences Research Council
(grant number BB/H008217/1), together with support from the UK Medical Research
Council and University of Cambridge, UK.References
[1] Cole, J.H., & Franke, K. (2017) Predicting Age Using
Neuroimaging: Innovative Brain Ageing Biomarkers. Trends Neurosci, 40:681-690.
[2] Bennett, I. J., & Madden, D. J. (2014). Disconnected aging:
cerebral white matter integrity and age-related differences in cognition.
Neuroscience, 276, 187-205.
[3] Alexander, A.L., Hurley, S.A., Samsonov, A.A., Adluru, N.,
Hosseinbor, A.P., Mossahebi, P., Tromp, D.P., Zakszewski, E., Field, A.S.
(2011) Characterization of cerebral white matter propertie using quantitative
magnetic resonance imaging stains. Brain connectivity, 1:423-446.
[4] Mwangi, B., Hasan, K.M., Soares, J.C. (2013) Prediction of
individual subject's age across the human lifespan using diffusion tensor
imaging: a machine learning approach. Neuroimage, 75:58-67.
[5] Mirzaalian, H., Ning, L., Savadjiev, P., Pasternak, O., Bouix, S.,
Michailovich, O., ... & Flashman, L. A. (2018). Multi-site harmonization of
diffusion MRI data in a registration framework. Brain imaging and behavior,
12(1), 284-295.
[6] Taylor, J.R., Williams, N., Cusack, R., Auer, T., Shafto, M.A.,
Dixon, M., Tyler, L.K., CamCAN, Henson, R.N. (2016). The Cambridge Centre for
Ageing and Neuroscience (CamCAN) data repository: Structural and functional
MRI, MEG, and cognitive data from a cross-sectional adult lifespan sample.
NeuroImage. doi: 10.1016/j.neuroimage.2015.09.018.
[7] Shafto, M.A., Tyler, L.K., Dixon, M., Taylor, J.R., Rowe, J.B.,
Cusack, R., Calder, A.J., Marslen-Wilson, W.D., Duncan, J., Dalgleish, T.,
Henson, R.N., Brayne, C., CamCAN, & Matthews, F.E. (2014). The Cambridge
Centre for Ageing and Neuroscience (CamCAN) study protocol: a cross-sectional,
lifespan, multidisciplinary examination of healthy cognitive ageing. BMC
Neurology, 14(204). doi:10.1186/s12883-014-0204-1.
[8] Hsu, Y.C., Tseng, W.Y. (2018) An
efficient regularization method for diffusion MAP-MRI estimation. 2018
ISMRM-ESMRMB Joint Annual Meeting.
[9] Chen, Y.J., Lo, Y.C., Hsu, Y.C.,
Fan, C.C., Hwang, T.J., Liu, C.M., Chien, Y.L., Hsieh, M.H., Liu, C.C., Hwu,
H.G., Tseng, W.Y. (2015) Automatic whole brain tract-based analysis using
predefined tracts in a diffusion spectrum imaging template and an accurate
registration strategy. Hum Brain Mapp, 36:3441-58.