Fitting a Directional Microstructure Model to Diusion-Relaxation MRI Data with Self-Supervised Machine Learning

2025-04-27 0 0 8.04MB 13 页 10玖币

侵权投诉

Fitting a Directional Microstructure Model to

Diﬀusion-Relaxation MRI Data with

Self-Supervised Machine Learning

Jason P. Lim1, Stefano B. Blumberg1,2, Neil Narayan1, Sean C. Epstein1,

Daniel C. Alexander1, Marco Palombo1,3,4, and Paddy J. Slator1

1Centre for Medical Image Computing, Department of Computer Science, University

College London, London, UK

2Centre for Artiﬁcial Intelligence, Department of Computer Science, University

College London, UK

3Cardiﬀ University Brain Research Imaging Centre (CUBRIC), School of

Psychology, Cardiﬀ University, Cardiﬀ, UK

4School of Computer Science and Informatics, Cardiﬀ University, Cardiﬀ, UK

p.slator@ucl.ac.uk

Abstract. Machine learning is a powerful approach for ﬁtting microstruc-

tural models to diﬀusion MRI data. Early machine learning microstruc-

ture imaging implementations trained regressors to estimate model pa-

rameters in a supervised way, using synthetic training data with known

ground truth. However, a drawback of this approach is that the choice

of training data impacts ﬁtted parameter values. Self-supervised learn-

ing is emerging as an attractive alternative to supervised learning in this

context. Thus far, both supervised and self-supervised learning have typ-

ically been applied to isotropic models, such as intravoxel incoherent mo-

tion (IVIM), as opposed to models where the directionality of anisotropic

structures is also estimated. In this paper, we demonstrate self-supervised

machine learning model ﬁtting for a directional microstructural model.

In particular, we ﬁt a combined T1-ball-stick model to the multidimen-

sional diﬀusion (MUDI) challenge diﬀusion-relaxation dataset. Our self-

supervised approach shows clear improvements in parameter estimation

and computational time, for both simulated and in-vivo brain data, com-

pared to standard non-linear least squares ﬁtting. Code for the artiﬁcial

neural net constructed for this study is available for public use from

the following GitHub repository: https://github.com/jplte/deep-T1-ball-

stick.

Keywords: Microstructure Imaging ·Machine Learning ·Self-supervised

learning

1 Introduction

Microstructure imaging aims to quantify features of the tissue microstructure

from in-vivo MRI [1]. Historically, microstructure imaging utilised diﬀusion MRI

arXiv:2210.02349v1 [eess.IV] 5 Oct 2022

2 J. Lim et al.

(dMRI) data. Recently, combined diﬀusion-relaxation MRI - where relaxation-

encoding parameters such as inversion time (TI) and echo time (TE) are varied

alongside diﬀusion-encoding parameters such as b-value and gradient direction

- has been emerging as an extension [23]. The typical approach to estimating

tissue microstructure from such diﬀusion or diﬀusion-relaxation data is multi-

compartment modelling, which utilises signal models comprising linear combi-

nations of multiple compartments - such as balls, sticks, zeppelins, and spheres

- each representing a distinct tissue geometry[21].

Multi-compartment microstructure models are usually ﬁt to the data with

non-linear least squares (NLLS) algorithms. However, these can be computa-

tionally expensive and are prone to local minima, necessitating grid searches or

parameter cascading [9] to seek global minima. Machine learning is a powerful

alternative. Thus far, most machine learning microstructure model ﬁtting ap-

proaches have used supervised learning [10,18,17,19,20,7,14]. However, a crucial

limitation is that the distribution of training data signiﬁcantly aﬀects ﬁtted pa-

rameters [12,8]. It has also proved diﬃcult to estimate directional parameters,

such as ﬁbre direction, with existing machine learning methods instead directly

estimating rotationally invariant parameters, such as mean diﬀusivity, fractional

anisotropy, mean kurtosis, and orientation dispersion. This may be due to the

diﬃculty of constructing a training dataset that adequately samples the high-

dimensional parameter space, and/or complications due to the periodicity of

angular parameters.

Self-supervised (sometimes imprecisely called unsupervised in the microstruc-

ture imaging context) learning is an alternative with the potential to address

these limitations. Self-supervised algorithms learn feature representations from

the input data by inferring supervisory constraints from the data itself. For mi-

crostructure imaging, self-supervised learning has been implemented with vox-

elwise fully connected artiﬁcial neural networks (ANNs). However, thus far self-

supervised microstructure imaging has been limited to isotropic models[11,6], in-

cluding many intravoxel incoherent motion (IVIM) MRI examples[2,13,25,26,8].

To our knowledge, self-supervised model ﬁtting has not yet been demonstrated

for directional microstructural models.

In this paper, we ﬁt an extended T1-ball-stick model to diﬀusion-relaxation

MRI data using self-supervised machine learning and demonstrate several ad-

vantages of this approach over classical NLLS, such as higher precision and faster

computational time.

2 Methods

2.1 Microstructure model

As this is a ﬁrst attempt at ﬁtting directional multi-compartment models with

self-supervised learning, we choose a simple model – the ball-stick model ﬁrst

proposed by Behrens et al. [3]. According to the ball-stick model, the expression

for the normalized signal decay is

S(b, g) = fexp −bλ||(g.n)+ (1 −f) exp (−bλiso) (1)

Self-supervised directional microstructure modelling 3

where bis the b-value, gis the gradient direction, λ|| and λiso are the parallel

and isotropic diﬀusivities of the stick and ball respectively, and nis the stick

orientation, which we parameterise using polar coordinates. The relationship be-

tween Cartesian and polar coordinates is n= [sin θcos φ, sin θsin φ, cos θ] where

φ∈[0, π] and θ∈[−π, π].

We extend the ball-stick model to account for T1 relaxation time, by as-

suming the ball and stick compartments have separate T1 times, represented

by T1ball and T1stick respectively. Note that we assume a single T2 for both

compartments, so the volume fraction fwill be aﬀected by the T2 of each

compartment. Given a combined T1 inversion recovery [5] and diﬀusion MRI

experiment, where inversion time (TI), b-value and gradient direction are simul-

taneously varied, we can ﬁt the following T1-ball-stick equation

S(b, g, TI, TR) = fexp −bλ||(g.n)



1−2 exp −TI

T1stick exp −TR

T1stick 



+ (1 −f) exp (−bλiso)



1−2 exp −TI

T1ball exp −TR

T1ball 



(2)

In this work, we ﬁrst ﬁt this model to combined T1-diﬀusion data with standard

NLLS, then demonstrate self-supervised ﬁtting with an ANN. We ﬁrst describe

the data, then the model ﬁtting techniques.

2.2 Combined T1-diﬀusion in vivo data

We utilise in-vivo data from 5 healthy volunteers (3 F, 2 M, age=19–46 years),

acquired from the 2019 multidimensional diﬀusion (MUDI) challenge [22]. The

acquisition sequence comprises simultaneous diﬀusion, inversion recovery (giving

T1 contrast), and multi-echo gradient echo (giving T2* contrast) measurements.

We chose to ignore the subsection of the data that is sensitive to T2* by only

included signals captured with the lowest echo time (80 ms). This is since the two

higher TEs have very low signal intensity and the 3 TEs have a small range, and

thus there is limited T2* information in the data. Our subsequent description

hence only refers to the subsection of the data with TE = 80 ms.

The datasets were obtained using a clinical 3T Philips Achieva scanner (Best,

Netherlands) with a 32-channel adult headcoil. Each scan includes 416 volumes

distributed over ﬁve b-shells, b ∈ {0,500,1000,2000,3000}s/mm2, with 16 uni-

formly spread directions, and 28 inversion times (TI) ∈[20, 7322] ms. For all

datasets, the following parameters were ﬁxed: repetition time TR=7.5 s, reso-

lution=2.5 mm isotropic, FOV=220×230×140 mm, SENSE=1.9, halfscan=0.7,

multiband factor 2, total acquisition time 52 min (including preparation time).

The MUDI data has already undergone standard pre-processing, see [22] for

full details. Upon inspection, we noted that the lowest (20 ms) and highest (7322

ms) inversion times, which comprise 7.14% of the data, were clearly dominated

by noise and/or artifacts (see Figure 1). We therefore removed them from the

data prior to model ﬁtting, leaving 416 MRI volumes in total. After removing

these TIs, the dataset contains 26 TIs ∈[176, 4673] ms.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

FittingaDirectionalMicrostructureModeltoDiusion-RelaxationMRIDatawithSelf-SupervisedMachineLearningJasonP.Lim1,StefanoB.Blumberg1;2,NeilNarayan1,SeanC.Epstein1,DanielC.Alexander1,MarcoPalombo1;3;4,andPaddyJ.Slator11CentreforMedicalImageComputing,DepartmentofComputerScience,UniversityCollegeLondon,L...

展开>> 收起<<

Fitting a Directional Microstructure Model to Diusion-Relaxation MRI Data with Self-Supervised Machine Learning.pdf

共13页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Fitting a Directional Microstructure Model to Diusion-Relaxation MRI Data with Self-Supervised Machine Learning

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: