diffusion Magnetic Resonance Image Analyzer (diMaRIA) page
Contents
- NEWS and related events
- Short Introduction
- Download
- Quick Reference
- Snapshots
- Techinical Details
- Known Limitations
and Future Solutions
- Acknowledgement
NEWS and related events
2024.09.20-21
- Invited lectures in JSMRM2024 synposiums
Masutani Y, Estimation of diffusion MRI model parameters by synthetic Q-space learning
Masutani Y, Introduction of R&D Environments for Synthetic Q-space Learning: An Instance
2024.09.09
-
Oral presentations in JSMRM2024
Masutani Y, Ichinoski Y, Konya K, Kato E, Mori N, Ota H, Mugikura S, and Takase K, DKI parameter inference by synthetic X-Q space learning and its application for breast dMRI
Ichinoski Y, Konya K, Mugikura S, Mori N, and Masutani Y, Estimation of NODDI model parameters in cohort data by synthetic Q-space learning: comparison with AMICO NODD
Konya K, Ichinoski Y, Kato E, Mori N, Ota H, Mugikura S, Takase K, and Masutani Y, Estimation of IVIM-DKI model parameters in clinical data of breast diffusion MRI by synthetic Q-space learning
2024.08.10 (virtual)
-
Hands-On Seminar for synthetic Q-space learning programming organized by JSMRM study group (material in Japanese)
2024.08.07
-
Oral presentation in JAMIT2024
Masutani Y, Parameter Inference for Diffusional Kurtosis Imaging by Synthetic X-Q Space Learning
2023.09.09
-
Oral presentation in JSMRM2023
Yamazaki K, Masutani Y, Uchida W, Sasaki K, Kamagata K, and Aoki S, Free water Imaging parameter estimation by combining synthetic Q-space learning and fitting: validation in white and grey matter
2023.09.09
-
Poster presentation in JSMRM2023
Yamazaki K, Masutani Y, Uchida W, Sasaki K, Kamagata K, and Aoki S, Free water Imaging parameter estimation by combining synthetic Q-space learning and fitting: validation in white and grey matter
2023.06.08
-
Poster presentation in ISMRM2023
Yamazaki K, Masutani Y, Uchida W, Kamagata K, Sasaki K, and Aoki S, Free water imaging parameter estimation by combination of synthetic q-space learning and conventional fitting: a hybrid approach
2022.09.09
-
Poster presentations in JSMRM2022
Sasaki K, and Masutani Y, Robust inference of diffusional kurtosis by combination of synthetic Q-space learning and DWI denoising
Yamazaki K, Masutani Y, Uchida W, Sasaki K, Kamagata K, Aoki S, Parameter inference by using synthetic Q-space learning in free water imaging
2022.09.01
-
The developer of diMaRIA: Yoshi MASUTANI moved to Tohoku University Medical School, Japan
2022.05.12
-
Poster presentation in ISMRM2022 (virtual)
Masutani Y, and Sasaki K, Single-Shell Free Water Imaging by Synthetic Q-Space Learning
2021.09.11
-
Poster presentation in JSMRM2021 (won the Research Encouragement Award)
Sasaki K, and Masutani Y, Comparison of Two Approaches for Diffusional Kurtosis Inference: Synthetic Q-Space Learning and DWI Denoising
2020.09.12
-
Poster presentation in JSMRM2020
Sasaki K, Fujiwara T and Masutani Y, On the Effect of Noise Mixture for Diffusional Kurtotis Inference by Synthetic Q-Space Learning
2019.09.20
-
Poster presentation in JSMRM2019
Fukunaga I, et al., Comparison of diMaRIA NODDI and AMCO NODDI - a study by two-sell dMRI data
2019.05.15
-
Poster presentation in ISMRM2019
Sasaki K, and Masutani Y, DKI parameter inference by deep neural networks trained by synthetic data
2019.04.08
-
Poster presentation in IEEE ISBI2019
Masutani Y, Noise Level Matching Improves Robustness of Diffusion MRI Parameter Inference by Synthetic Q-Space Learning
2019.03.13
-
diMaRIA19b was released.
Short Introduction
This software: diffusion Magnetic Resonance Image Analyzer (diMaRIA) is a plugin of a general purpose volume data display software: VOLUME-ONE. It is a successor of dTV-II software for DTI analysis and visualization plugin for VOLUME-ONE. The current version of diMaRIA has following functions.
(1) DTI model based tractography
- simple e1 (principal direction of diffusion tensor) tracking or anisotropic decomposition (similar to tensorline)
(2) Computation of parameter maps by standard DTI model
-
fractional anisotpry (FA), color FA, mean diffusivity (MD or ADC), and isotropic diffusion
(3) Extra parameter maps computed by deep regression neural networks
- DKI model: axial/radial diffusion coefficient (D), axial/radial diffusional kurtosis (K)
- NODDI model: fraction of isotropic diffusion (Fiso), fraction of intra cellular diffusion (Fic), and orientation dispersion (OD)
(4) diffusion profile display
-
tensor ellipsoids,
ADC profile, etc.
(5) statistical analysis (mean and std. dev.) within ROI
- DTI model: FA and MD (ADC)
- DKI model:
axial/radial D and K
- NODDI model:
Fiso, Fic and OD
Note: parameters of (3) and (5) are for current version of diMaRIA. Future extension may add other models and parameters.
Downolad
Licence Agreement for VOLUME-ONE and diMaRIA
Read following items carefully before downloading the software tools.
1. All Rights Reserved
The developpper (Yoshitaka MASUTANI) possesses all rights including copyright and ownership of the software tools. Any modification of the software including language localization, and software analysis including reverse engineering, decompose, and disassemble are prohibited.
2. No Guarantee
Those software toolis are distributed in the hope of the developper that it will be useful, but WITHOUT WARRANTY, without even the implied warranty of MERCHANTABILITY or FITNESS FOR PARTICULAR PURPOSE. The developper does NOT have responsibilities for any damages on material and immaterial properties of the users and third parties.
3. Academic Research Purpose Only
This software is only for academic research. Please do not use for clinical or commercial purpose including promotion of commercial products.
4. In Disclosure of Analysis Results
In disclosure of analysis results using the software tools at any media such as journal paper, webpage, and reports, please add description to the articles, such as "This result was obtained by using the free software diMaRIA developed by Y. Masutani.", and refer the papers below according to the employed technical features.
- DTI basic functions: tractography and parameter maps and statitical analysis of FA and ADC, etc. -> [1]
- radial/axial DKI
parameter computation -> {2], [3]
- NODDI analysis ->[3]
[1]
Masutani@Y, et al., MR diffusion tensor imaging: recent advance and new techniques for diffusion tensor visualization, EJR:46(1), pp.53-66, 2003
[2] Masutani Y, et al.,
Toward Analytic Computation of Fiber-Radial Diffusional Kurtosis by Q-Space Data Representation with Radial Basis Functions, proc. ISMRM, 2017
[3] Masutani Y, et al.,
Noise level matching improves robustness in diffusion MRI parameter inference by synthetic Q-space learning, proc. IEEE ISBI ,2019
5. Questions and Contact
Basically, no more information beyond description in this web site is provided.
Download
Currently, ONLY WINDOWS VERSION IS AVAILABLE. The sample data set of a healthy volunteer is by courtesy of Hiroshima Heiwa Clinic.
Download VOLUME-ONE and diMaRIA (a ZIP file of a folder including all components) from here (7.21 MB).
-
"Volume-One1.93.exe" and other binary components (DLLs and standard plugins)
- "diMaRIA19b.exe" in "plugins\tools\diMaRIA" and "diMaRIa.rtnnFile.bin" (neural network parameters) in "plugins\tools\diMaRIA\rtnn_data"
- readme.txt
Download sample data (a ZIP file of a folder including raw data, MPG text file and readme) from here (78.2 MB).
- sample_dMRI.raw
- sample_MPG.txt
- readme.txt
Quick Reference
If you are familiar to the former software: dTV-II, you can use diMaRIA similarly. If you are unfamiliar to dTV-II, please follow the steps below.
How To Start
Step 1: data preparation
You need to prepare two types of data. See also the readme file attached to the sample data for analysis of your data.
(1) DWI data as multi-channel volume data
(2) MPG data as text data corresponding to the DWI
Raw format of concatnated DWI volume data is recommended. In the current version, only single channel of b=0 DWI (So) is allowed. If you have more, average them or leave only one at the first channel.
Step 2: Launching software tools and loading data
You need to launch two software tools.
(1) First, launch VOLUME-ONE software by double-clicking the icon.
(2) Load DWI data to
VOLUME-ONE as raw data (or analyze format, etc.) by the "File->Import->Raw" command and specify the dimension (matrix size, etc.) of volume data.
(3) If needed, adjust the data by mirroring, rotation, interpolation, byte-swap, ... so that you can correctly observe your DWI volume data of isotropic resolution.
(4) To launch diMaRIA, select the menu of VOLUME-ONE; "Tools->Launch Plug-in..." and choose the file "tools\diMaRIA\diMaRIA19b.exe" in the file selection dialog.
(5) Load MPG data to diMaRIA
from the menu "File->Import->MPG text file".
Step 3: Use of main functions
(1) Tractography
- Set ROI(s) in the palettes and press "Analyze and Display" button, and the result is displayed in VOLUME-ONE. (see Fig.1 in shapshots)
(2) Parameter maps
- You can
get parameter maps by the following step(s), and get the results in VOLUME-ONE.
-- "Special->Compute standard DTI parameter maps" for FA, etc.
-- "Special->Extra maps by Run Time Neural Network" for other parameters of DKI and NODDI models, and select the map(s) to be computed in the preview window, and then press "Compute" button.
(see Fig.2)
(3) Statistical Analysis
- Set ROI(s) in the palette and press "Analyze and Display" button, and the results are displayed in the text palette. You can select and copy the result text there, and can export to other software by pasting it. (see Fig.3)
Others
- Realtime display of parameters at a voxel pointed by VOLUME-ONE cursor: "Special->Realtime analysis ON-OFF" (see Fig.4)
- Save DWI data in volume-one format: "FIle->Save", and you do not need to specify the attributes of volume data when loading again.
- ROI editor to create multiple ROI(s) in the list of ROI pallette.
-
Voxelizer of tractography results (may be useful for Tract-Specific Analysis): "Special->Voxelize visible trackline as->Voxelgroup ROI"
- ROI object(s) display in VOLUME-ONE: "Setting->ROI 3D Object->Show, Update (Hide)"
- and features inheritated from dTV-II
Snapshots
Below, you can find the main features of diMaRIA and VOLUME-ONE visually.
Fig.1 Overview (diMaRIA works with VOLUME-ONE)
Fig.2 Preview for extra parameter maps by deep neural networks
Fig.3 Statistical analysis within ROI
Fig.4 Realtime analysis for single location
Techinical Details
diMaRIA software is based on the following technical features.
(1) Q-space learning
The effectiveness of Q-space learning to infer dMRI parameters instead of conventional model fitting is reported by Golkov, et al. [1]. They employed the multi-layer perceptron (MLP), and diMaRIA also uses MLP. Certain part of the configurations and training parameters of MLP in diMaRIA were determined by referring their work.
(2) Synthetic Q-space learning
Masutani [2] and Ye [3] showed feasibility and usefulness of
synthetic Q-space learning, that is, training by only synthetic data. In addition, Masutani [4] experimantally showed that the level of noise added to training data is one of the keys for robust dMRI parameter inference. The current MLP in diMaRIA was trained at very standard noise level of dMRI. MLPs trained at other noise levels are also planned to be provided.
(3) Q-space/B-space representation by radial basis functions
A practical issue of dMRI parameter inference by Q-space learning is that the location of input values in Q-space must be identical to those at training. There is no guarantee for such situation in most cases for analyzing your data. For solving this issue, diMaRIA employs the radial basis functions (RBF), which is an efficient interpolator/extrapolator to reconstruct smooth function from non-grid data [5]. After reconstruction of signal decay function in Q-space for the data to be analyzed, diMaRIA resamples at the Q-space locations identical to those of training data. Actually, based on preliminary experiments, interpolation results between in Q-space and in B-space have no siginificant difference. Therefore, diMaRIA performs B-space interpolation/resampling by RBF. This technique is also used for reconstruct 1D smooth series of fiber-radial signal decays for robust inference of radial kurtosis [6].
(4) Run-Time Neural Network (RTNN)
Though the training of neural network for diMARIA was intensively performed in a high performance comupting environment with GPU, diMaRIA
uses only CPU (but, with multi-thread and SIMD) when analyzing your data. That is, diMaRIA has no training facility inside but has functions for only prediction of dMRI parameters. This modiule is called as "Run-Time Neural Network (RTNN)" in diMaRIA. The training result is in a binary format, and is loaded automatically when diMaRIA is launched.
[1] Golkov V, et al., Q-space deep learning: twelve-fold shorter and model-free diffusion MRI scans, IEEE TMI, 35(5):1344–1351, Apr. 2016.
[2] Masutani Y, et al., Synthesis and Extension of training data for diffusion MRI parameter inference by deep regression, proc. 2nd MAIAMI, Tokyo, Jun. 2018 (in Japanese)
[3] Ye C, et al., Q-space learning with synthesized training data, Proc. CDMRI 2018, Granada, Sep. 2018
[4] Masutani Y, Noise level matching improves robustness in diffusion MRI parameter inference by synthetic Q-space learning, proc. IEEE ISBI, Venice, Apr. 2019
[5] Masutani Y, et al., Unstructured sampling and RBF-based ODF reconstruction in Q-space for diffusion MR tractography, proc CARS2015, Barcelona, Jun. 2015
[6]
Masutani Y, et al.,
Toward Analytic Computation of Fiber-Radial Diffusional Kurtosis by Q-Space Data Representation with Radial Basis Functions, proc. ISMRM, Honolulu, Apr. 2017
[7] Masutani Y, Recent Advances in Parameter Inference for Diffusion MRI Signal Models Magn Reson Med Sci, 21(1):132-147, 2022
Known Limitations and Future Solutions
Currently, several limitations and problems below are recognized by the developper, and will be improved in the near future.
(1) File format for DWI and MPG
- NIfTI, or other standard format should be covered (offers for converter or codes for it are very welcome and appreciated!)
(2) Single b=0
- Auto preprocessor is planned to be implemented to average multi So channels
(3) Noise level matching
- Preparation of multiple neural networks
trained with various noise levels and automatic/semi-sutomatic selection of the optimal network are planned to be implemented
(4) "RTNN file NOT found" error
- If you launch VOLUME-ONE by a shortcut, default RTNN file is NOT loaded correctly when launching diMaRIA and you need to load it manually from the menu "File->Import->RTNN file".
Acknowledgement
This work was financially supported by two research projects below. The developper (Yoshi Masutani) are grateful to all the members of the projects.
- MEXT KAKENHI JP26108003 "Multidisciplinary Computational Anatomy" (FY2014-FY2018)
-
JST CREST RESEARCH GRANT JPMJCR15D1 "New Challenges for Mathematical Modeling in Clinical Medicine" (FY2015-FY2020)
- JSPS KAKENHI 21K12650 "synthetic X-Q space learning" (FY2021-FY2023)
Also,
I would like to thank Prof. Daniel C. Alexander, Dr. Gary Hui Zhang, Mr. Ryu Tanno and the members of CMIC/UCL for fruitful discussion and hospitality during my short visit to their laboratory.