Y T.Benchekroun1, D Christiaens1, A Schuh2, A Makropoulos2, T Poppe1, E Hughes1, J Hutter1, A Price1, J-D Tournier1, L Cordero-Grande1, S Counsell1, D Rueckert2, J Hajnal1, D Batalle1,3, M Deprez1, AD Edwards1,4*
1 Centre for the Developing Brain, King’s College London, London SE1 7EH
2 Department of Computing, Imperial College London, Exhibition Rd, London SW7 2AZ
3 Department of Forensic and Neurodevelopmental Science, King’s College London, London SE5 8AF
4 MRC Centre for Neurodevelopmental Disorders, King’s College London, London SE11 1YR
Introduction (include hypothesis)
During the perinatal period, human brain functional and structural development underpins the expansion of cognitive abilities. Diffusion MRI provides key insights into the evolution of brain networks during this period . Nevertheless, identification of the complex patterns in such data requires advanced analysis methods. We hypothesised that structural connectomes derived from diffusion MR data would encode rich information about brain development, and we tested whether this could be extracted using Deep Learning, taking age at birth and age at scan as exemplars.
Methods (include source of funding and ethical approval if required)
From the developing Human Connectome Project (dHCP) data set, structural connectivity data was available for 524 neonates born between 23 and 42 weeks of gestational age (GA) and scanned between 37 and 45 weeks post menstrual age (PMA). Infants were scanned during natural sleep and T2-weighted and diffusion MRI scans were motion corrected . Each infant’s structural connectivity was constructed by calculating the SIFT2-weighted sum of streamlines  connecting each pair of the 90 regions segmented from T2w volumes. Balanced random sampling was used to stratify the dataset in a training (N=418, 80%) and an independent testing set (N=106, 20%). Vectorised structural connectivity matrices of the training set were used as input to train a fully connected Neural Network predicting PMA and GA (3 hidden layers for PMA prediction and 6 hidden layers for GA prediction). Performance of the predictive algorithm was tested in the independent testing set.
We reached a Mean Absolute Error (MAE) of 0.65 weeks for prediction of PMA at scan in the testing group. Correlation between true and predicted PMA was r = 0.93 (p<0.001).
We reached a MAE of 1.5 weeks for prediction of GA at birth in the testing group. Correlation between true and predicted age was r = 0.85 (p<0.001).
Deep Learning methods can predict age at birth and age at scan with high accuracy from the neonatal structural connectome. The neonatal structural connectome therefore provides important information about neurodevelopmental maturation and can facilitate studies of typical and atypical neonatal brain development. The dHCP data used for this study will be made available to researchers shortly.
References (include acknowledgement here if appropriate)
 Batalle, D., et al., Early development of structural networks and the impact of prematurity on brain connectivity. Neuroimage, 2017.
 Christiaens, D., et al., Learning compact q-space representations for multi-shell diffusion-weighted MRI. IEEE Trans Med Imaging, 2018.
 Smith, R.E., et al., SIFT2: Enabling dense quantitative assessment of brain white matter connectivity using streamlines tractography. Neuroimage, 2015.