Introduction

Structural connectome analysis using diffusion tractography helps quantify and lend insights to brain network abnormalities associated with neurological disorders^1,2. 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.

Methods

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 C_i=1,2,…,M, consisting of a series of streamlines $$$f^j$$$ connecting both k^th and l^th regions, where M is the total number of C_i existing in S{k,l}. For each C_i, 70%/30% of total streamlines were used to construct the training/test set. For each $$$f^j$$$ of the training C_i, 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 $$$f^j$$$ ∈ R^3x100 denotes the deep feature and $$$c^j$$$ ∈ R^3x100 denotes a centroid of the streamlines in C_i. After training, the fully connected layer produced the output probability vector, P(C_i=1,2,…,M|$$$f^j$$$) = softmax(w·(G($$$f^j$$$)◦r) + b) where P(C_i|$$$f^j$$$) is the prediction probability of the input $$$f^j$$$ belonging to the class C_i, w is the convolution filter, G($$$f^j$$$) is the output of max pooling layer, “◦” is the element-wise multiplication operator, r ∈ R^m 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(C_i=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 C_i=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}: C_i, the trained DCNN classified a given streamline $$$f^j$$$ into a “reproducible streamline” if its prediction probability P(C_i|$$$f^j$$$) was greater than β×max(P(C_i|$$$f^j$$$)), where β is a fractional threshold controlling the likelihood of reference streamline in C_i (e.g., β=0 yields no effect of DCNN prediction and selects all possible streamlines in C_i, while, β=1 selects the single streamline having the highest DCNN prediction likelihood of reference streamline in C_i). Finally, intraclass correlation coefficient (ICC)⁷ was assessed to measure the reproducibility of S{k,l} as a function of β.

Results

1477 streamline classes C_i 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 subject_1-3 due to the exclusion of the many outliers marked by white arrows.

Discussion

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.

Conclusion

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.

Acknowledgements

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