Improving reproducibility of diffusion connectome analysis using deep convolutional neural network model
Min-Hee Lee1, Nolan Baird O'Hara2, and Jeong-Won Jeong3

1Pediatrics and Translational Imaging Laboratory, Wayne State University School of Medicine, Detroit, MI, United States, 2Translational Neuroscience Program, Wayne State University School of Medicine, Detroit, MI, United States, 3Pediatrics, Neurology and Translational Imaging Laboratory, Wayne State University School of Medicine, Detroit, MI, United States


Reproducibility of diffusion-weighted structural connectomes is highly dependent on acquisition and tractography model, limiting the interpretation of connectomes acquired in the clinical setting. This study proposes a novel deep convolutional neural network (DCNN) to improve the reproducibility of structural connectomes, by which highly reproducible streamlines can be identified via an end-to-end deep learning of reference streamline coordinates in Human Connectome Project diffusion data. Preliminary results demonstrate that the proposed DCNN prediction model can improve the reproducibility of clinical connectomes (31.29% of F-statistics in intraclass correlation coefficient) and effectively remove noisy streamlines based on based on their poor prediction probabilities.


Structural connectome analysis using diffusion tractography helps quantify and lend insights to brain network abnormalities associated with neurological disorders1,2. However, its reproducibility is highly dependent on acquisition and tractography model, often limiting the interpretation of structural connectomes in the clinical setting3. This study proposes a novel deep convolutional neural network (DCNN)4 to improve the reproducibility of structural connectomes by showing that highly reproducible streamlines (i.e., consistently present across subjects) can be identified via an end-to-end deep learning of reference streamline coordinates derived from high quality human connectome project (HCP) diffusion data. We assume that such a DCNN-based streamline prediction model could provide a unique tool to effectively remove “wiggly tracked” streamlines based on their improbability in the trained DCNN, and improve the reproducibility of individual structural connectomes for clinical applications.


To define reference streamlines of interest, this study utilized the HCP 3T diffusion MRI data pool (https://db.humanconnectome.org/). Briefly, the MRtrix software package (http://www.mrtrix.org/) was initially used to generate a fiber orientation distribution (FOD)5 group template. From this template, 50 million tracts were generated by applying SIFT1 reconstruction to 100 million iFOD2-ACT whole brain tracts6. The AAL parcellation atlas (http://www.gin.cnrs.fr/en/tools/aal-aal2/) was then used to create a whole brain connectome, S{k,l}, in which each element defines a reference streamline class Ci=1,2,…,M, consisting of a series of streamlines $$$f^j$$$ connecting both kth and lth regions, where M is the total number of Ci existing in S{k,l}. For each Ci, 70%/30% of total streamlines were used to construct the training/test set. For each $$$f^j$$$ of the training Ci, our DCNN model (Fig. 1) was designed to learn 3-D (x,y,z) coordinates of 100 equal-number streamline segments by minimizing center loss, $$$L_c^j=\frac{1}{2}\parallel{f^{j}}-{c^{j}}\parallel_2^2$$$, where $$$f^j$$$ ∈ R3x100 denotes the deep feature and $$$c^j$$$ ∈ R3x100 denotes a centroid of the streamlines in Ci. After training, the fully connected layer produced the output probability vector, P(Ci=1,2,…,M|$$$f^j$$$) = softmax(w·(G($$$f^j$$$)◦r) + b) where P(Ci|$$$f^j$$$) is the prediction probability of the input $$$f^j$$$ belonging to the class Ci, w is the convolution filter, G($$$f^j$$$) is the output of max pooling layer, “◦” is the element-wise multiplication operator, r ∈ Rm is the dropout mask vector of Bernoulli variables with probability 0.5 of being set as 0, and b ∈ R is the bias term. An argument of the maximum P(Ci=1,2,…,M|$$$f^j$$$) was used to predict class membership of the input $$$f^j$$$. F1 score was utilized to evaluate overall performance of correct prediction across Ci=1,2,..,M. For validation, 13 healthy controls were scanned with a 3T Siemens Verio using 64 encoding directions at three b-values: 1000, 2000 and 3000 s/mm2. The same connectome procedure used for HCP data was applied to individual diffusion data, generating S{k,l} in HCP template space. In each class of S{k,l}: Ci, the trained DCNN classified a given streamline $$$f^j$$$ into a “reproducible streamline” if its prediction probability P(Ci|$$$f^j$$$) was greater than β×max(P(Ci|$$$f^j$$$)), where β is a fractional threshold controlling the likelihood of reference streamline in Ci (e.g., β=0 yields no effect of DCNN prediction and selects all possible streamlines in Ci, while, β=1 selects the single streamline having the highest DCNN prediction likelihood of reference streamline in Ci). Finally, intraclass correlation coefficient (ICC)7 was assessed to measure the reproducibility of S{k,l} as a function of β.


1477 streamline classes Ci were selected from S{k,l} of HCP (i.e., M=1477 having streamline count >1000). After 80 epochs, center loss of our DCNN model was converged at 0.0146/0.011, yielding a high average macro F1 score of 0.951/0.956 in training/test set. In all validation data of b=1000/2000/3000 s/mm2, our DCNN model could increase 31.29%/31.18%/27.30% of F-statistics in ICC value at β=0.8 (Fig. 2, ICC =0.9129/0.8959/0.8893 without DCNN prediction and ICC=0.9323/0.9188/0.9111 with DCNN prediction). As demonstrated in Fig. 3, this improvement could be attributed to effective removal of “wiggly tracked” outliers, based on their low DCNN-determined prediction probability, P(Ci|$$$f^j$$$). For instance, it is clear that the DCNN-determined C1159 (streamlines in S{1,77}) has better consistency across subject1-3 due to the exclusion of the many outliers marked by white arrows.


This study translates deep learning techniques to improve the reproducibility of individual diffusion connectomes. Our preliminary data show that DCNN-based streamline prediction can accomplish this by controlling a single experimental parameter: fractional threshold of maximal DCNN prediction probability. Further investigation is needed to determine the effectiveness of DCNN prediction in analyzing more sophisticated network properties.


Our findings provide preliminary evidence supporting the utility of end-to-end deep learning of HCP white matter trajectories as a tool to improve the reproducibility of individual connectomes in a clinical setting.


This study was funded by a grant from the National Institute of Health, (R01-NS089659 to J.J).


  1. Jeong JW, Sundaram S, Behen ME, Chugani HT. Differentiation of speech delay and global developmental delay in children using DTI tractography-based connectome. Am J. Neuroradiol. 2016; 37:1170-7.
  2. Jeong JW, Asano E, Juhász C, Behen ME, Chugani HT. Postoperative axonal changes in the contralateral hemisphere in children with medically refractory epilepsy: A longitudinal diffusion tensor imaging connectome analysis. Hum Brain Mapp. 2016; 37:3946-56.
  3. Prčkovska V, Rodrigues P, Puigdellivol Sanchez A, Ramos M, Andorra M, Martinez-Heras E, Falcon C, Prats-Galino A, Villoslada P. Reproducibility of the Structural Connectome Reconstruction across Diffusion Methods. J Neuroimaging. 2016;26:46-57.
  4. HE K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. p. 770-778.
  5. Tournier JD, Calamante F, Connelly A. Robust determination of the fibre orientation distribution in diffusion MRI: non-negativity constrained super-resolved spherical deconvolution. Neuroimage, 2007;35:1459–72.
  6. Smith RE, Tournier JD, Calamante F, ConnellyA. SIFT: Spherical-deconvolution informed filtering of tractograms. NeuroImage, 2013;67:298-312.
  7. Telesford QK, Morgan AR, Hayasaka S, Simpson SL, Barret W, Kraft RA, Mozolic JL, Laurienti PJ. Reproducibility of graph metrics in FMRI networks. Front Neuroinform. 2010;4:117.


Network architecture of the proposed DCNN model, which learns 3-D coordinates of reference streamlines (input: $$$f^j$$$) and their ground truth classes (output: Ci) as defined in HCP group template space.

ICC of the DCNN-determined S{k,l} as a function of β. Note that β=0 indicates no DCNN was involved in predicting the likelihood of reference streamlines in S{k,l}.

Left: Examples of the DCNN-determined connectivity matrices S{k,l} showing improved reproducibility across subject1-3 (b=1000s/mm2). ICC(subject1,subject2)=0.9459/0.9610, ICC(subject1,subject3)=0.9271/0.9467, and ICC(subject2, subject3)=0.9178/0.9379, were obtained without/with DCNN prediction (β=0/0.2). Right: To illustrate the effect of DCNN prediction on individual links of S{k,l}, S{1,77} connecting left precentral gyrus (magenta colored) and left thalamus (cyan colored) was selected for streamlines in C1159 and analyzed without/with DCNN prediction (β=0/0.2). It is apparent that the DCNN with β=0.2 could improve the reproducibility of C1159 across subject1-3 by excluding many outliers (marked by white arrows) based on their low DCNN-determined prediction probability, P(C1159|$$$f^j$$$).

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)