A novel groupwise diffusion MRI registration framework using deep learning
Poster No:
2367
Submission Type:
Abstract Submission
Authors:
Junyi Wang1, Zhu Xi1, Mubai Du1, William Wells III2, Lauren O'Donnell2, Fan Zhang1
Institutions:
1University of Electronic Science and Technology of China, Chengdu, Sichuan, 2Harvard Medical School, Boston, MA
First Author:
Co-Author(s):
Introduction:
Diffusion MRI (dMRI)1 enables the estimation of the underlying brain tissue microstructure and tracking of white matter (WM) fiber tracts in vivo2. Groupwise registration of dMRI data from multiple individuals is an important task for brain template construction3 to identify normative brain structures and their variability across different populations for potential diagnosis and treatment of diseases4,5. However, groupwise dMRI registration is a challenging task due to the unique nature of dMRI data, which describes not only the strength but also the orientation of water diffusion in brain tissues. Widely used methods perform registration using dMRI-derived scalar images (e.g., fractional anisotropy; FA) across individuals and neglect the important orientation information that can reveal the underlying WM fiber orientation.
We propose a novel deep-learning approach for simultaneous registration of dMRI data across multiple individuals into a common space. This work is built on our previous DDMReg6 method for pairwise dMRI registration and further enables groupwise dMRI registration for brain template construction. Compared with a baseline and a state-of-the-art method, we show a significant improvement of fiber spatial alignment after registration, while maintaining a good volumetric overlap of segmented brain tissue structures.
We propose a novel deep-learning approach for simultaneous registration of dMRI data across multiple individuals into a common space. This work is built on our previous DDMReg6 method for pairwise dMRI registration and further enables groupwise dMRI registration for brain template construction. Compared with a baseline and a state-of-the-art method, we show a significant improvement of fiber spatial alignment after registration, while maintaining a good volumetric overlap of segmented brain tissue structures.
Methods:
From the dMRI data of each subject, we compute two types of input data: 1) FA, which is useful for aligning anatomical structures from the whole brain, and 2) 16 tract orientation maps (TOMs)7, which are useful for aligning streamlines of specific anatomical WM tracts. The underlying network architecture is based on the popular conditional deformable template creation method8. To handle the large amount of input data, we use a two-step training process, including 1) pre-training a subnetwork on each TOM to compute tract-specific templates, and 2) overall network training on FA with the pre-trained tract-specific subnetworks as backbones. After learning, the network learns a deformable field for each subject to warp the FA and TOMs of all subjects into the common template space. In addition, we introduce bidirectional learning that allows reversed registration from the learned template to each subject. Our loss function is shown in Fig 1.
For the experiment, we utilize dMRI data from 100 subjects from the Human Connectome Project (HCP)9 (20 for training, 80 for testing). FA images are computed using SlicerDMRI10 and TOMs are computed using TractSeg7. We compare our method with two template creation methods including the widely used SyN algorithm and the deep learning-based VoxelMorph method. The evaluation metric includes tract distance to assess fiber spatial alignment of the WM tracts, and tissue Dice for volume overlap of the tissue segmentations (i.e., gray matter, white matter, and cerebrospinal fluid).
For the experiment, we utilize dMRI data from 100 subjects from the Human Connectome Project (HCP)9 (20 for training, 80 for testing). FA images are computed using SlicerDMRI10 and TOMs are computed using TractSeg7. We compare our method with two template creation methods including the widely used SyN algorithm and the deep learning-based VoxelMorph method. The evaluation metric includes tract distance to assess fiber spatial alignment of the WM tracts, and tissue Dice for volume overlap of the tissue segmentations (i.e., gray matter, white matter, and cerebrospinal fluid).
Results:
Fig 2a gives visualization results of the three comparison methods. Our method obtains the warped TOMs that can better align the baseline image, showing a better groupwise registration across subjects. Fig 2b shows the fiber spatial alignment evaluation results. For each tract, there is a significant difference across the 3 methods (via an ANOVA test), where the proposed method obtains overall the best performance. Fig 2c shows the results of the tissue segmentation overlap, where the VoxelMorph and proposed methods generate significantly better performance than the baseline SyN method.
Conclusions:
We develop a deep learning framework that effectively combines tissue orientation and brain structure information for groupwise dMRI registration. We demonstrate advanced registration performance compared to existing groupwise dMRI registration methods. The proposed method can serve as an important tool for building brain atlases.
Modeling and Analysis Methods:
Diffusion MRI Modeling and Analysis
Image Registration and Computational Anatomy 2
Novel Imaging Acquisition Methods:
Diffusion MRI 1
Keywords:
Data Registration
Machine Learning
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review
Provide references using author date format
2. Zhang, F. et al. (2022). Quantitative mapping of the brain’s structural connectivity using diffusion MRI tractography: A review. NeuroImage, 249, 118870.
3. Zhang, H. et al. (2007). Unbiased white matter atlas construction using diffusion tensor images. Medical Image Computing and Computer-Assisted Intervention: MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention, 10(Pt 2), 211–218.
4. Zhang, F. et al. (2018). An anatomically curated fiber clustering white matter atlas for consistent white matter tract parcellation across the lifespan. NeuroImage, 179, 429–447.
5. Yeh, F.-C. et al.(2018). Population-averaged atlas of the macroscale human structural connectome and its network topology. NeuroImage, 178, 57–68.
6. Zhang, F. et al. (2022). Deep Diffusion MRI Registration (DDMReg): A Deep Learning Method for Diffusion MRI Registration. IEEE Transactions on Medical Imaging, 41(6), 1454–1467.
7. Wasserthal, J.et al. (2019). Combined tract segmentation and orientation mapping for bundle-specific tractography. In arXiv [cs.CV]. arXiv. http://arxiv.org/abs/1901.10271
8. Dalca, A. et al. (2019). Learning conditional deformable templates with convolutional networks. Advances in Neural Information Processing Systems, 32. https://proceedings.neurips.cc/paper/2019/hash/bbcbff5c1f1ded46c25d28119a85c6c2-Abstract.html
9. Elam, J. S. et al. (2021). The Human Connectome Project: A retrospective. NeuroImage, 244, 118543.
10. Zhang, F. et al. (2020). SlicerDMRI: Diffusion MRI and Tractography Research Software for Brain Cancer Surgery Planning and Visualization. JCO Clinical Cancer Informatics, 4, 299–309.