Rui Zeng1, Jinglei Lv2, He Wang3, Luping Zhou2, Michael Barnett2, Fernando Calamante2, and Chenyu Wang2
1School of Biomedical Engineering, The University of Sydney, Sydney, Australia, 2The University of Sydney, Sydney, Australia, 3Fudan University, Shanghai, China
Synopsis
In
this study, a deep learning model called FODSRM was developed for fiber
orientation distribution (FOD) super-resolution, which enhances single-shell
low-angular-resolution FOD computed from clinic-quality dMRI data (e.g., 32
directions b=1000) to obtain the super-resolved high-angular resolution quality
that would have been produced from advanced research scanners (e.g.,
multi-shell HARDI data). The results demonstrate that the super-resolved FOD
data generated by the proposed method can generate high-definition structural
connectome from clinical acquisition protocols, even when applied to data from
a protocol not included in the trained dataset.
Introduction
Mapping
the human connectome permits the study of brain connectivity and yields new
insights into neuroscience1.
However, reliable connectome reconstruction using diffusion MRI (dMRI) data
acquired by widely available low-angular resolution clinical protocols remains
a challenge, which further limits connectome applications at larger clinical
scale.Purpose
We
developed a deep-learning-based framework for fiber orientation distribution
(FOD) super-resolution (FODSR). Our method enhances single-shell
low-angular-resolution (LAR) FOD computed from clinic-quality dMRI data to
obtain the super-resolved high-angular resolution quality that would have been
produced from advanced research scanners. The super-solved FODs can be used for
reliable connectome reconstruction, which has close quality to that obtained
from multi-shell high-angular-resolution state-of-the-art research dMRI protocols.Method
Our deep learning model FODSRM (fiber
orientation distribution super-resolution method) takes (sequentially) as input
an FOD patch cropped from a single-shell LAR FOD image, which is generated by
the state-of-the-art single-shell 3-tissue constrained spherical deconvolution
(SS3T CSD)2, and outputs the
super-resolved version of the central voxel of this patch. Once the
super-resolved FOD voxel for each patch has been obtained, they are combined
together in the reverse order of the cropping operation to generate the
super-resolved FOD data, which is the output of FODSRM. Then the super-resolved FOD
data is used for connectome reconstruction, which is performed by SIFT2-based3 whole-brain tractography4, and then the Desikan-Killiany atlas was
employed to calculate the connectome. (see Figure 1 for the pipeline). Human Connectome
Project (HCP) data5 were selected to train our
model as it is one of the largest dMRI dataset acquired with sophisticated
protocols. Furthermore, as single-shell 32-direction dMRI images with low b
value (b=1000) are widely used in practical clinical settings, we generated
clinically accessible dMRI data (i.e. single-shell LAR) from
the original multi-shell high-angular-resolution: we subsampled the first 32
directions from the b=1000 s/mm2 shell, as the HCP protocol samples
directions in an incremental way such that any sequential subset would result
in an approximately optimal design. Finally, the b0 volume is added into
the extracted 32 directions to obtain the complete single-shell low-angular
resolution dMRI dataset. The ‘ground truth’ FOD data used for conducting
evaluation are generated from the original multi-shell HCP dMRI data using
multi-shell CSD6. 110 subjects from HCP dataset are randomly
selected for conducting our experiments. Our dataset is split into 50 training 10 evaluation, and 10 validation. For assessment, we compared FODSRM predictions with the ground
truth and with the single-shell LAR FOD images computed from the current state-of-the-art
method (SS3T CSD). Furthermore, we generated
connectomes using each method and compared the ground truth with FODSRM and
with SS3T CSD.Results
Figure 1 shows three anatomical
regions, which are zoomed to reveal FODs. FODSRM was applied to single-shell
low-angular-resolution FOD data of test subjects previously unseen by the
network during training; the features of crossing, kissing, and fanning are
clearly resolved by FODSRM, and the super-resolved FOD images have very good
agreement with the ground truth FOD data. To quantitatively measure the
improvement brought by FODSRM, we reported the angular correlation coefficients7 between the ground truth and
the FOD estimates (see Figure 3) in three different types of tissues, such as white
matter (WM), voxels of WM partial volumed with subcortical grey matter (SGM), and
with cortical grey matter (CGM). One can clearly see that our method
outperformed SS3T CSD. Furthermore, mean angular error (MAE), which is computed
from the angular discrepancy between estimated and ground truth fixel8 (a fiber bundle instance
extracted from a FOD), is used to evaluate our FOD estimates. Figure 4 demonstrates
the MAE results in WM, WM&CGM, and WM&SGM, where we can see FODSRM
outperformed SS3T CSD. Figure 5 shows the average connectomes (generated from
SS3T CSD, FODSRM, and MSMT CSD respectively) across 10 subjects in HCP dataset
and the difference between the ground truth with SS3T CSD and with FODSRM. We
can observe that FODSRM yielded the most accurate connectome reconstructions,
outperforming SS3T CSD by a large margin. To test
the generalization ability, FDOSRM was applied to clinical diffusion MR
images acquired by unseen (not present in the training phase) protocols using a
different MRI scanner. From Figure 5, we can find that the improvement of
connectome reconstruction obtained from super-resolved FOD images of LARDI data
is very consistent with the HCP results.Discussion
The
proposed deep-learning-based method, FODSRM, allows for the generation of
super-resolved FOD images directly from conventional clinical-type single-shell
LAR dMRI data, leading to comparable quality to those FODs produced from more
advanced multi-shell HARDI data and superior results that current
state-of-the-art deconvolution methods, Importantly, this is achieved with only
a fraction of the data (e.g. as little as ~11% of the dMRI data for the HCP
data used here). Furthermore, FODSRM achieves these FOD improvements without
requiring high b-value data, which is most often not available in clinical
protocols, but has been shown to be required for optimal performance when using
current FOD algorithms (MSMT CSD).Acknowledgements
No acknowledgement found.References
1. Fornito, A., Zalesky, A. & Bullmore,
E. Fundamentals of Brain Network Analysis. (Elsevier, 2016).
doi:10.1016/C2012-0-06036-X
2. Khan, W. et al. Three-tissue
compositional analysis reveals in-vivo microstructural heterogeneity of white
matter hyperintensities following stroke. Neuroimage 218, 116869
(2020).
3. Smith, R. E., Tournier, J.-D.,
Calamante, F. & Connelly, A. SIFT2: Enabling dense quantitative assessment
of brain white matter connectivity using streamlines tractography. Neuroimage
119, 338–351 (2015).
4. Smith, R. E., Tournier, J.-D.,
Calamante, F. & Connelly, A. Anatomically-constrained tractography:
Improved diffusion MRI streamlines tractography through effective use of
anatomical information. Neuroimage 62, 1924–1938 (2012).
5. Van Essen, D. C. et al. The
WU-Minn Human Connectome Project: An overview. Neuroimage 80, 62–79
(2013).
6. Jeurissen, B., Tournier, J.-D. D.,
Dhollander, T., Connelly, A. & Sijbers, J. Multi-tissue constrained
spherical deconvolution for improved analysis of multi-shell diffusion MRI
data. Neuroimage 103, 411–426 (2014).
7. Anderson, A. W. Measurement of fiber
orientation distributions using high angular resolution diffusion imaging. Magn.
Reson. Med. 54, 1194–1206 (2005).
8. Raffelt, D. et al. Apparent Fibre
Density: A novel measure for the analysis of diffusion-weighted magnetic
resonance images. Neuroimage 59, 3976–3994 (2012).