Synthesizing DWI with high b-values from low b-value datasets using deep learning
Poster No:
2247
Submission Type:
Abstract Submission
Authors:
Hsin-Tzu Huang1, Li-Wei Kuo2, Chun-Hung Yeh3, Yi-Ping Chao1
Institutions:
1Department of Computer Science and Information Engineering, Chang Gung University, Taoyuan, Taiwan, 2Institute of Biomedical Engineering and Nanomedicine, National Health Research Institutes, Miaoli, Taiwan, 3Department of Medical Imaging and Radiological Sciences, Chang Gung University, Taoyuan, Taiwan
First Author:
Hsin-Tzu Huang
Department of Computer Science and Information Engineering, Chang Gung University
Taoyuan, Taiwan
Department of Computer Science and Information Engineering, Chang Gung University
Taoyuan, Taiwan
Co-Author(s):
Li-Wei Kuo
Institute of Biomedical Engineering and Nanomedicine, National Health Research Institutes
Miaoli, Taiwan
Institute of Biomedical Engineering and Nanomedicine, National Health Research Institutes
Miaoli, Taiwan
Chun-Hung Yeh
Department of Medical Imaging and Radiological Sciences, Chang Gung University
Taoyuan, Taiwan
Department of Medical Imaging and Radiological Sciences, Chang Gung University
Taoyuan, Taiwan
Yi-Ping Chao
Department of Computer Science and Information Engineering, Chang Gung University
Taoyuan, Taiwan
Department of Computer Science and Information Engineering, Chang Gung University
Taoyuan, Taiwan
Introduction:
Diffusion MRI (dMRI) has been widely applied in various fields. In comparison to the clinically common diffusion tensor imaging (DTI), high angular resolution diffusion imaging (HARDI) possesses the capability to resolve fiber crossings and provides more information about white matter neuroanatomy, such as fiber density. However, for various HARDI techniques, multi-shell diffusion weighted images (DWI) with higher b-values are required, leading to an increased imaging time. This may render it less suitable for clinical use. To address this issue, a previous study proposed computed DWI, synthesizing arbitrary b-value DWI from a set of measured b-value images through voxel-wise fitting. In this study, deep learning techniques were employed using a model combining DenseNet and Generative Adversarial Network (GAN) to establish a model with low b-value DWI as input and high b-value DWI as output. The feasibility of the model was assessed by comparing the ground truth with the generated DWI, generalized Q-sampling image (GQI) indices mapping derived from generated DWI, and fiber density map derived from generated DWI.
Methods:
All dMRI data were collected with a GE MR 750 scanner at Chang Gung Memorial Hospital, Linkou, Taiwan. Two-shell DWI with b-values of 1000 and 2000 s/mm², along with 60 non-collinear diffusion gradient directions on each shell, were conducted in 115 subjects. The data were partitioned into training, validation, and testing sets, with 89, 10, and 16 subjects, respectively. A multi-level densely connected network with GAN was employed as the data synthesis model. Paired DWI with b-value of 1000 s/mm2 (DWI-b1000) and 2000 s/mm2 (DWI-b2000) with the same diffusion direction were used as input and output for training, undergoing 200 epochs. For the evaluation of the model's feasibility, structural similarity index (SSIM), peak signal-to-noise ratio (PSNR) and mean squared error (MSE) were assessed between DWI-b1000, generated DWI-b2000 and ground truth DWI-b2000. Furthermore, the (GQI) indices mapping derived from generated DWI, and fiber density map derived from generated DWI were also assessed in comparing with DWI-b1000 only and ground truth DWI-b2000. The DSI Studio software and MRtrix 3 were utilized for GQI reconstruction and estimation of voxel-wise total fiber density.
·Figure 1 The flowchart of data processing. (a) data spliting and deep learning. (b) performance evaluation.
Results:
The results show that the SSIM/PSNR/MSE between the original DWI-b1000 and the ground truth (GT) DWI-b2000 are 0.82/35.68/0.00092, while the SSIM/PSNR/MSE between the generated DWI-b2000 and the GT DWI-b2000 are 0.862/36.83/0.0006. The differences between the two are statistically significant (p<0.001). This indicates that the generated DWI-b2000 is indeed more similar to the GT DWI-b2000 in image quality assessment.
In GQI indices mapping, the SSIM/PSNR/MSE between the quantitative anisotropy (QA) estimation using original DWI-b1000 only and using original DWI-b1000 and GT DWI b-2000 are 0.97/57.57/0.00039 and the SSIM/PSNR/MSE between the QA estimation using original DWI-b1000 and generated DWI b-2000 and using original DWI-b1000 and GT DWI b-2000 are 0.965/57.11/0.00045. The mean angle error (the formula is shown in Figure 1) between using original DWI-b1000 only and using original DWI-b1000 and GT DWI b-2000, and using original DWI-b1000 and generated DWI b-2000 and using original DWI-b1000 and GT DWI b-2000 are 11.09 and 11.2, respectively.
Based on voxel-wise total fiber density, in comparison with using only DWI-b1000 (SSIM/PSNR/MSE=0.985/60.75/0.00014), the estimation using DWI-b1000 and generated DWI-b2000 is significantly more similar to the ground truth data (SSIM/PSNR/MSE=0.998/65.79/0.000023) (p<0.001).
In GQI indices mapping, the SSIM/PSNR/MSE between the quantitative anisotropy (QA) estimation using original DWI-b1000 only and using original DWI-b1000 and GT DWI b-2000 are 0.97/57.57/0.00039 and the SSIM/PSNR/MSE between the QA estimation using original DWI-b1000 and generated DWI b-2000 and using original DWI-b1000 and GT DWI b-2000 are 0.965/57.11/0.00045. The mean angle error (the formula is shown in Figure 1) between using original DWI-b1000 only and using original DWI-b1000 and GT DWI b-2000, and using original DWI-b1000 and generated DWI b-2000 and using original DWI-b1000 and GT DWI b-2000 are 11.09 and 11.2, respectively.
Based on voxel-wise total fiber density, in comparison with using only DWI-b1000 (SSIM/PSNR/MSE=0.985/60.75/0.00014), the estimation using DWI-b1000 and generated DWI-b2000 is significantly more similar to the ground truth data (SSIM/PSNR/MSE=0.998/65.79/0.000023) (p<0.001).
·Figure 2 The results of image quality assessment
Conclusions:
This study developed a deep learning-based method for generating high b-value DWI, and the generated images are quite similar to the ground truth images. In comparison to assessing fiber density using only lower b-value DWI, the inclusion of generated higher b-value DWI demonstrates significant improvement in effectiveness.
Modeling and Analysis Methods:
Methods Development 2
Neuroinformatics and Data Sharing:
Informatics Other 1
Keywords:
Informatics
Machine Learning
MRI
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
Other - deep learning
1|2Indicates the priority used for review
Provide references using author date format
Boulanger, M., (2021). Deep learning methods to generate synthetic CT from MRI in radiotherapy: A literature review. Physica Medica, 89, 265-281. https://doi.org/https://doi.org/10.1016/j.ejmp.2021.07.027
Higaki, T., (2018). Introduction to the Technical Aspects of Computed Diffusion-weighted Imaging for Radiologists. Radiographics, 38(4), 1131-1144. https://doi.org/10.1148/rg.2018170115
Wakana, S., (2005). Macroscopic orientation component analysis of brain white matter and thalamus based on diffusion tensor imaging. Magnetic Resonance in Medicine, 53(3), 649-657. https://doi.org/10.1002/mrm.20386
Xie, S., (2015). How does B-value affect HARDI reconstruction using clinical diffusion MRI data? PLoS One, 10(3), e0120773. https://doi.org/10.1371/journal.pone.0120773
Zhong, Z., (2023). Accelerating High b-Value Diffusion-Weighted MRI Using a Convolutional Recurrent Neural Network (CRNN-DWI). Bioengineering, 10(7), 864. https://www.mdpi.com/2306-5354/10/7/864