Jointly Learning in x-q space for Fiber ODF Estimation
Poster No:
2355
Submission Type:
Abstract Submission
Authors:
Chengdong Deng1, Haotian Jiang1, Jaeil Kim2, Jiquan Ma#1, Geng Chen#3
Institutions:
1Department of Computer Science and Technology, Heilongjiang University, Harbin, China, 2School of Computer Science and Engineering, Kyungpook National University, Daegu, South Korea, 3School of Computer Science and Engineering, Northwestern Polytechnical University, Xi'an, China
First Author:
Co-Author(s):
Jaeil Kim
School of Computer Science and Engineering, Kyungpook National University
Daegu, South Korea
School of Computer Science and Engineering, Kyungpook National University
Daegu, South Korea
Geng Chen#
School of Computer Science and Engineering, Northwestern Polytechnical University
Xi'an, China
School of Computer Science and Engineering, Northwestern Polytechnical University
Xi'an, China
Introduction:
Fiber orientation distribution functions (fODFs) provide vital information for reconstructing the 3D geometric structure of brain white matter. Existing methods to predict fODFs are confronted by two inherent limitations. Firstly, the traditional methods necessitate densely sampled diffusion MRI (dMRI) data in q-space. Secondly, the deep learning methods neglect a comprehensive consideration of the x-q space information [1-3].
To address the aforementioned limitations, we propose a novel fiber orientation estimation method, called MXQ, which consists of an x-space learning model based on a 3D U-Net [4] and a q-space learning model based on spherical convolutional neural networks (SCNNs) [6-7]. Instead of considering a single domain, our joint learning framework exploits and fuses x-q space information to predict fODFs. The fusion is achieved by a simple and effective integration module that can transform the features from two domains. Extensive experiments demonstrate that our MXQ model is effective in fODF estimation and outperforms cutting-edge models remarkably.
To address the aforementioned limitations, we propose a novel fiber orientation estimation method, called MXQ, which consists of an x-space learning model based on a 3D U-Net [4] and a q-space learning model based on spherical convolutional neural networks (SCNNs) [6-7]. Instead of considering a single domain, our joint learning framework exploits and fuses x-q space information to predict fODFs. The fusion is achieved by a simple and effective integration module that can transform the features from two domains. Extensive experiments demonstrate that our MXQ model is effective in fODF estimation and outperforms cutting-edge models remarkably.
Methods:
Our MXQ aims to improve the joint learning of x-space and q-space information through the use of a hybrid architecture consisting of a 3D U-Net and SCNNs, as illustrated in Figure 1 (a).
The spatial relationship of each voxel in x-space is learned by the 3D U-Net, which is well-suited for capturing spatial correlation information, i.e.,
$$
F_{\mathrm{x}} = \text{MX}(A_{\mathrm{shc}} ),
$$
where MX represents our x-space module, which is used to learn the features of x-space Fx, and Ashc represents the input represented by spherical harmonic coefficients (SHCs).
We use SCNNs to learn q-space information by mapping fODFs to spherical data, i.e.,
$$
F_{\mathrm{q}} = \text{MQ}(A_{\mathrm{amp}} ),
$$
where MQ represents our q-space module used to learn the features of q-space Fq, and Aamp denotes the spherical data transformed from Ashc.
In SIF, as shown in Figure 1 (b), we fuse the x-space and q-space features by mining the commonality and individuality of two kinds of features with the channel attention components. Fianlly, the SIF module predicts the fODFs Yxq via
$$
Y_{\mathrm{xq}} = \text{SIF}(F_{\mathrm{x}}, F_{\mathrm{q}})
$$
The spatial relationship of each voxel in x-space is learned by the 3D U-Net, which is well-suited for capturing spatial correlation information, i.e.,
$$
F_{\mathrm{x}} = \text{MX}(A_{\mathrm{shc}} ),
$$
where MX represents our x-space module, which is used to learn the features of x-space Fx, and Ashc represents the input represented by spherical harmonic coefficients (SHCs).
We use SCNNs to learn q-space information by mapping fODFs to spherical data, i.e.,
$$
F_{\mathrm{q}} = \text{MQ}(A_{\mathrm{amp}} ),
$$
where MQ represents our q-space module used to learn the features of q-space Fq, and Aamp denotes the spherical data transformed from Ashc.
In SIF, as shown in Figure 1 (b), we fuse the x-space and q-space features by mining the commonality and individuality of two kinds of features with the channel attention components. Fianlly, the SIF module predicts the fODFs Yxq via
$$
Y_{\mathrm{xq}} = \text{SIF}(F_{\mathrm{x}}, F_{\mathrm{q}})
$$
·Figure 1: An Overview of our framework.
Results:
Dataset: The dataset used in our experiments is from the human connectome project (HCP) [5]. We employed MSMT-CSD and the entire 270 gradient directions to compute the ground truth fODFs. To emulate data acquisition in a clinical setting, we downsampled the dMRI data by uniformly selecting 30 gradient directions at b = 1000 s/mm2. For the MX module, we set the size of input patches to 16 × 16 × 16. For the MQ module, we transformed the 45 SHCs into spherical data with 42 vertices.
Experimental Results: As shown in Figure 2 (a) and (b), our MXQ provides the best ACC and MAE results. Figure 2 (c), indicate that our MXQ provides the best visualization results that are close to the ground truth. Exiting methods are capable of accurately predicting single-direction fODFs. However, when dealing with crossing fODFs, particularly those with three crossings, our MXQ demonstrates superior accuracy in comparison with existing methods.
Experimental Results: As shown in Figure 2 (a) and (b), our MXQ provides the best ACC and MAE results. Figure 2 (c), indicate that our MXQ provides the best visualization results that are close to the ground truth. Exiting methods are capable of accurately predicting single-direction fODFs. However, when dealing with crossing fODFs, particularly those with three crossings, our MXQ demonstrates superior accuracy in comparison with existing methods.
·Figure 2: Quantitive and qualitative evaluation results.
Conclusions:
We proposed MXQ, a deep learning-based fODF estimation model that exploits and fuses the x-q space features. Experimental results show that our MXQ outperforms existing models in terms of both quantitative and qualitative evaluations.
Modeling and Analysis Methods:
Diffusion MRI Modeling and Analysis 2
Other Methods
Novel Imaging Acquisition Methods:
Diffusion MRI 1
Keywords:
MRI
Other - Fiber ODF
1|2Indicates the priority used for review
Provide references using author date format
[2] Lin, Z. (2019), 'Fast learning of fiber orientation distribution function for MR tractography using convolutional neural network', Medical Physics, 46(7): 3101–3116.
[3] Lucena, O. (2021), 'Enhancing the estimation of fiber orientation distributions using convolutional neural networks', Computers in Biology and Medicine, 135: 104643.
[4] Ronneberger, O. (2015). 'U-Net: Convolutional networks for biomedical image segmentation'. In MICCAI 2015, Munich, Germany, October 5-9, 2015, Proceedings, Part III18, 234–241. Springer.
[5] Van Essen, D. C. (2012), 'The Human Connectome Project: a data acquisition perspective'. NeuroImage, 62(4): 2222–2231.
[6] Zhao, F. (2019). 'Spherical U-Net on cortical surfaces: methods and applications'. In IPMI 2019, Hong Kong, China, June 2–7, 2019, Proceedings 26, 855–866. Springer.
[7] Zhao, F. (2021), 'Spherical deformable U-Net: Application to cortical surface parcellation and development prediction'. IEEE Transactions on Medical Imaging, 40(4): 1217–1228.