Convolutional Neural Networks for Fiber Orientation Distribution Enhancement to Improve Single-Shell Diffusion MRI Tractography

Oeslle Lucena; Sjoerd B. Vos; Vejay Vakharia; John Duncan; Sebastien Ourselin; Rachel Sparks

doi:10.1007/978-3-030-52893-5_9

Convolutional Neural Networks for Fiber Orientation Distribution Enhancement to Improve Single-Shell Diffusion MRI Tractography

Oeslle Lucena^*, Sjoerd B. Vos, Vejay Vakharia, John Duncan, Sebastien Ourselin, Rachel Sparks

^*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceeding › Chapter › peer-review

2 Citations (Scopus)

Abstract

Diffusion MRI (dMRI) tractography may help locate critical white matter (WM) tracts that should be preserved during neurosurgery. A key step in this process is estimating fiber orientation distribution (FOD), often done from a model such as constrained spherical deconvolution (CSD). Multi-shell (MS) multi-tissue CSD (M-CSD) provides a robust WM FOD by estimating the relative contribution to the dMRI signal from each tissue type (WM, grey matter, and cerebrospinal fluid), however, single-shell (SS) single tissue CSD (S-CSD) cannot independently estimate the signal contribution for each tissue type. S-CSD is therefore less accurate estimating FOD in voxels where multiple tissues are present. Due to that inaccuracy, tractography using S-CSD often generates more spurious WM streamlines compared to M-CSD. In this work, we present a framework to regress the M-CSD model coefficients from the S-CSD model coefficients using a convolutional neural network (CNN) in order to improve tractography. We construct a training dataset comprising acquired MS dMRI and paired synthetic SS dMRI, generated by selecting the outer shell from the MS dMRI. We select a High Resolution Network (HighResNet) as our choice of CNN to ensure subtle details of the CSD models are preserved during regression. The HighResNet is trained to perform patch-based regression from the S-CSD model coefficients and a co-registered T1-wieghted MR (T1) to the M-CSD model coefficients. We evaluate the method on patients with epilepsy who appeared structurally normal on T1. Four WM tracts related to language are extracted using a ROI-based probabilistic tractography. For comparison, M-CSD is as a pseudo ground truth. The original S-CSD generated tracts with Dice of 0.53–0.64, and the HighResNet regressed CSD models generated tracts with Dice of 0.73–0.77. We demonstrate HighResNet can regress M-CSD model coefficients from S-CSD model coefficients resulting in tracts more similar to the M-CSD generated tracts and with fewer spurious streamlines than S-CSD generated tracts.

Original language	English
Title of host publication	Mathematics and Visualization
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	101-112
Number of pages	12
DOIs	https://doi.org/10.1007/978-3-030-52893-5_9
Publication status	Published - 2020

Publication series

Name	Mathematics and Visualization
ISSN (Print)	1612-3786
ISSN (Electronic)	2197-666X

Keywords

Deep learning
Diffusion weighted image
Tractography

Access to Document

10.1007/978-3-030-52893-5_9

Cite this

Lucena, O., Vos, S. B., Vakharia, V., Duncan, J., Ourselin, S., & Sparks, R. (2020). Convolutional Neural Networks for Fiber Orientation Distribution Enhancement to Improve Single-Shell Diffusion MRI Tractography. In Mathematics and Visualization (pp. 101-112). (Mathematics and Visualization). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-030-52893-5_9

@inbook{1d0828a7235d4bf1afd2d8071718f907,

title = "Convolutional Neural Networks for Fiber Orientation Distribution Enhancement to Improve Single-Shell Diffusion MRI Tractography",

abstract = "Diffusion MRI (dMRI) tractography may help locate critical white matter (WM) tracts that should be preserved during neurosurgery. A key step in this process is estimating fiber orientation distribution (FOD), often done from a model such as constrained spherical deconvolution (CSD). Multi-shell (MS) multi-tissue CSD (M-CSD) provides a robust WM FOD by estimating the relative contribution to the dMRI signal from each tissue type (WM, grey matter, and cerebrospinal fluid), however, single-shell (SS) single tissue CSD (S-CSD) cannot independently estimate the signal contribution for each tissue type. S-CSD is therefore less accurate estimating FOD in voxels where multiple tissues are present. Due to that inaccuracy, tractography using S-CSD often generates more spurious WM streamlines compared to M-CSD. In this work, we present a framework to regress the M-CSD model coefficients from the S-CSD model coefficients using a convolutional neural network (CNN) in order to improve tractography. We construct a training dataset comprising acquired MS dMRI and paired synthetic SS dMRI, generated by selecting the outer shell from the MS dMRI. We select a High Resolution Network (HighResNet) as our choice of CNN to ensure subtle details of the CSD models are preserved during regression. The HighResNet is trained to perform patch-based regression from the S-CSD model coefficients and a co-registered T1-wieghted MR (T1) to the M-CSD model coefficients. We evaluate the method on patients with epilepsy who appeared structurally normal on T1. Four WM tracts related to language are extracted using a ROI-based probabilistic tractography. For comparison, M-CSD is as a pseudo ground truth. The original S-CSD generated tracts with Dice of 0.53–0.64, and the HighResNet regressed CSD models generated tracts with Dice of 0.73–0.77. We demonstrate HighResNet can regress M-CSD model coefficients from S-CSD model coefficients resulting in tracts more similar to the M-CSD generated tracts and with fewer spurious streamlines than S-CSD generated tracts.",

keywords = "Deep learning, Diffusion weighted image, Tractography",

author = "Oeslle Lucena and Vos, {Sjoerd B.} and Vejay Vakharia and John Duncan and Sebastien Ourselin and Rachel Sparks",

year = "2020",

doi = "10.1007/978-3-030-52893-5_9",

language = "English",

series = "Mathematics and Visualization",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "101--112",

booktitle = "Mathematics and Visualization",

address = "Germany",

}

Lucena, O, Vos, SB, Vakharia, V, Duncan, J, Ourselin, S & Sparks, R 2020, Convolutional Neural Networks for Fiber Orientation Distribution Enhancement to Improve Single-Shell Diffusion MRI Tractography. in Mathematics and Visualization. Mathematics and Visualization, Springer Science and Business Media Deutschland GmbH, pp. 101-112. https://doi.org/10.1007/978-3-030-52893-5_9

Convolutional Neural Networks for Fiber Orientation Distribution Enhancement to Improve Single-Shell Diffusion MRI Tractography. / Lucena, Oeslle; Vos, Sjoerd B.; Vakharia, Vejay et al.
Mathematics and Visualization. Springer Science and Business Media Deutschland GmbH, 2020. p. 101-112 (Mathematics and Visualization).

Research output: Chapter in Book/Report/Conference proceeding › Chapter › peer-review

TY - CHAP

T1 - Convolutional Neural Networks for Fiber Orientation Distribution Enhancement to Improve Single-Shell Diffusion MRI Tractography

AU - Lucena, Oeslle

AU - Vos, Sjoerd B.

AU - Vakharia, Vejay

AU - Duncan, John

AU - Ourselin, Sebastien

AU - Sparks, Rachel

PY - 2020

Y1 - 2020

N2 - Diffusion MRI (dMRI) tractography may help locate critical white matter (WM) tracts that should be preserved during neurosurgery. A key step in this process is estimating fiber orientation distribution (FOD), often done from a model such as constrained spherical deconvolution (CSD). Multi-shell (MS) multi-tissue CSD (M-CSD) provides a robust WM FOD by estimating the relative contribution to the dMRI signal from each tissue type (WM, grey matter, and cerebrospinal fluid), however, single-shell (SS) single tissue CSD (S-CSD) cannot independently estimate the signal contribution for each tissue type. S-CSD is therefore less accurate estimating FOD in voxels where multiple tissues are present. Due to that inaccuracy, tractography using S-CSD often generates more spurious WM streamlines compared to M-CSD. In this work, we present a framework to regress the M-CSD model coefficients from the S-CSD model coefficients using a convolutional neural network (CNN) in order to improve tractography. We construct a training dataset comprising acquired MS dMRI and paired synthetic SS dMRI, generated by selecting the outer shell from the MS dMRI. We select a High Resolution Network (HighResNet) as our choice of CNN to ensure subtle details of the CSD models are preserved during regression. The HighResNet is trained to perform patch-based regression from the S-CSD model coefficients and a co-registered T1-wieghted MR (T1) to the M-CSD model coefficients. We evaluate the method on patients with epilepsy who appeared structurally normal on T1. Four WM tracts related to language are extracted using a ROI-based probabilistic tractography. For comparison, M-CSD is as a pseudo ground truth. The original S-CSD generated tracts with Dice of 0.53–0.64, and the HighResNet regressed CSD models generated tracts with Dice of 0.73–0.77. We demonstrate HighResNet can regress M-CSD model coefficients from S-CSD model coefficients resulting in tracts more similar to the M-CSD generated tracts and with fewer spurious streamlines than S-CSD generated tracts.

AB - Diffusion MRI (dMRI) tractography may help locate critical white matter (WM) tracts that should be preserved during neurosurgery. A key step in this process is estimating fiber orientation distribution (FOD), often done from a model such as constrained spherical deconvolution (CSD). Multi-shell (MS) multi-tissue CSD (M-CSD) provides a robust WM FOD by estimating the relative contribution to the dMRI signal from each tissue type (WM, grey matter, and cerebrospinal fluid), however, single-shell (SS) single tissue CSD (S-CSD) cannot independently estimate the signal contribution for each tissue type. S-CSD is therefore less accurate estimating FOD in voxels where multiple tissues are present. Due to that inaccuracy, tractography using S-CSD often generates more spurious WM streamlines compared to M-CSD. In this work, we present a framework to regress the M-CSD model coefficients from the S-CSD model coefficients using a convolutional neural network (CNN) in order to improve tractography. We construct a training dataset comprising acquired MS dMRI and paired synthetic SS dMRI, generated by selecting the outer shell from the MS dMRI. We select a High Resolution Network (HighResNet) as our choice of CNN to ensure subtle details of the CSD models are preserved during regression. The HighResNet is trained to perform patch-based regression from the S-CSD model coefficients and a co-registered T1-wieghted MR (T1) to the M-CSD model coefficients. We evaluate the method on patients with epilepsy who appeared structurally normal on T1. Four WM tracts related to language are extracted using a ROI-based probabilistic tractography. For comparison, M-CSD is as a pseudo ground truth. The original S-CSD generated tracts with Dice of 0.53–0.64, and the HighResNet regressed CSD models generated tracts with Dice of 0.73–0.77. We demonstrate HighResNet can regress M-CSD model coefficients from S-CSD model coefficients resulting in tracts more similar to the M-CSD generated tracts and with fewer spurious streamlines than S-CSD generated tracts.

KW - Deep learning

KW - Diffusion weighted image

KW - Tractography

UR - http://www.scopus.com/inward/record.url?scp=85095865838&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-52893-5_9

DO - 10.1007/978-3-030-52893-5_9

M3 - Chapter

AN - SCOPUS:85095865838

T3 - Mathematics and Visualization

SP - 101

EP - 112

BT - Mathematics and Visualization

PB - Springer Science and Business Media Deutschland GmbH

ER -