Alexandra Cristobal Huerta

his thesis proposes and investigates novel techniques to shorten the acquisition of three-
T dimensional high-resolution brain and knee MR images, without deteriorating image
quality. The Three-dimensional Fast Spin Echo (3D-FSE) pulse sequence was chosen as
sequence to investigate since it allows depicting small structures of the body isotropically
in the three orthogonal planes of acquisition, and is especially relevant in clinical settings
for acquiring Proton Density (PD)-weighted and T 2 -weighted images from knee and brain.
However, its long acquisition time and its high Specific Absorption Rate (SAR) limit its
wider acceptance as standard sequence within clinical protocols.
An introduction to Magnetic Resonance (MR) physics is provided in Chapter 1. We
introduced the general physical concepts, the two main sequences on which most of the
current sequences are based on and the two advanced sequences investigated in this thesis,
with its advantages and disadvantages. We also explain the three most popular reconstruction
techniques currently available in most of the sequences and clinical scanners. We finalize this
chapter by introducing the concept of Deep Learning (DL) and its advantages.
In Chapter 2 we provided insights and guidance about the most suitable acceleration
technique among the ones available in clinical scanners to acquire faster high-resolution
PD-weighted knee images with a 3D-FSE sequence. We compared Half Fourier (HF), Parallel
Imaging (PI) and Compressed Sensing (CS) quantitatively in terms of Signal to Noise
Ratio (SNR), Contrast to Noise Ratio (CNR) and scanning time in a phantom and in-vivo
experiments. From our results we concluded that PI obtains the highest image quality among
the techniques but has in the longest acquisition time. When comparing HF versus CS, the
latest obtains more blurry images between tissues than HF for the same acquisition time.
Therefore, HF is proposed as the most suitable acceleration technique to use for PD-weighted
knee images with 3D-FSE in clinical scanners.
To further accelerate the acquisition of 3D-FSE sequences and reduce SAR, in Chapter 3
and Chapter 4 the Three-dimensional Gradient and Spin Echo (3D-GRASE) sequence was
implemented and investigated together with PI and CS for high resolution imaging. In these
chapters, several suitable cartesian k-space trajectories and k-space grids were proposed and
evaluated. Specifically, Chapter 3 evaluates four different k-space trajectories for obtaining
knee PD-weighted images and two k-space trajectories for brain T 2 -weighted images. These
trajectories were based on the SORT phase-encoding strategy combined with linear or radial
modulation. The performance of each trajectory was evaluated in simulation, in phantom and
in-vivo experiments by measuring the Point Spread Function (PSF) and Signal to Noise Ratio

his thesis proposes and investigates novel techniques to shorten the acquisition of three-
T dimensional high-resolution brain and knee MR images, without deteriorating image
quality. The Three-dimensional Fast Spin Echo (3D-FSE) pulse sequence was chosen as
sequence to investigate since it allows depicting small structures of the body isotropically
in the three orthogonal planes of acquisition, and is especially relevant in clinical settings
for acquiring Proton Density (PD)-weighted and T 2 -weighted images from knee and brain.
However, its long acquisition time and its high Specific Absorption Rate (SAR) limit its
wider acceptance as standard sequence within clinical protocols.
An introduction to Magnetic Resonance (MR) physics is provided in Chapter 1. We
introduced the general physical concepts, the two main sequences on which most of the
current sequences are based on and the two advanced sequences investigated in this thesis,
with its advantages and disadvantages. We also explain the three most popular reconstruction
techniques currently available in most of the sequences and clinical scanners. We finalize this
chapter by introducing the concept of Deep Learning (DL) and its advantages.
In Chapter 2 we provided insights and guidance about the most suitable acceleration
technique among the ones available in clinical scanners to acquire faster high-resolution
PD-weighted knee images with a 3D-FSE sequence. We compared Half Fourier (HF), Parallel
Imaging (PI) and Compressed Sensing (CS) quantitatively in terms of Signal to Noise
Ratio (SNR), Contrast to Noise Ratio (CNR) and scanning time in a phantom and in-vivo
experiments. From our results we concluded that PI obtains the highest image quality among
the techniques but has in the longest acquisition time. When comparing HF versus CS, the
latest obtains more blurry images between tissues than HF for the same acquisition time.
Therefore, HF is proposed as the most suitable acceleration technique to use for PD-weighted
knee images with 3D-FSE in clinical scanners.
To further accelerate the acquisition of 3D-FSE sequences and reduce SAR, in Chapter 3
and Chapter 4 the Three-dimensional Gradient and Spin Echo (3D-GRASE) sequence was
implemented and investigated together with PI and CS for high resolution imaging. In these
chapters, several suitable cartesian k-space trajectories and k-space grids were proposed and
evaluated. Specifically, Chapter 3 evaluates four different k-space trajectories for obtaining
knee PD-weighted images and two k-space trajectories for brain T 2 -weighted images. These
trajectories were based on the SORT phase-encoding strategy combined with linear or radial
modulation. The performance of each trajectory was evaluated in simulation, in phantom and
in-vivo experiments by measuring the Point Spread Function (PSF) and Signal to Noise Ratio

108 Summary

(SNR), and compared with similar 3D-FSE acquisitions. From this work it was concluded that
SORT Linear modulation encoding for T 2 -weighted images and SORT Radial modulation
encoding with M=0 for PD-weighted images obtain image quality comparable to 3D-FSE,
while reducing SAR by more than 40% and shortening acquisition time by 20%. On the
other hand, Chapter 4 investigates the effects of the different k-space trajectories proposed
in Chapter 3 together with two common k-space undersampling grids for CS combined with
PI (CSPI) in PD-weighted and T 2 -weighted 3D-GRASE acquisitions. CSPI requires an
incoherent undersampling, a variable density k-space grid and a fully sampled k-space center
in order to achieve an artefact-free reconstruction. Two undersampled grids proposed in
the literature for different sequences and applications fulfill these requirements: Variable
Density (VD) pseudo-random Gaussian grid and VD Poisson-disc grid. The incoherence
of the different combinations of k-space trajectories and undersampled k-space grids was
evaluated in simulation, phantom and in-vivo experiments, concluding that i) sampling
patterns combining a VD Poisson-disc k-space grid in both PD-weighted and T 2 -weighted
contrasts obtained the highest incoherence and ii) the trajectory has low influence on the
results.

Chapter 3 and Chapter 4 shown that the quality of 3D-GRASE images highly depends
on the trajectory applied during the acquisition. Moreover, the image contrast can slightly
change with respect to Fast Spin Echo (FSE) if the acquisition parameters are not carefully
chosen, due to the T -weighted contrast introduced by the Gradient Recalled Echos (GREs).
∗

In order to propose a solution to this problem, in Chapter 5 a Deep Learning (DL) method
that brings the appearance of 3D-GRASE closer to 3D-FSE images, removing artefacts and
achieving a more similar FSE image contrast, was investigated. Three different DL models
were developed based on a Three-dimensional (3D) U-Net in combination with three loss
functions previously proposed in the literature for regression problems: i) the voxel-wise
metric l 2 -norm, ii) Destructural Similarity Index (DSSIM) and iii) the perceptual loss. The
results from this work showed that the overall image quality in the axial plane is improved
when a 3D U-Net with a perceptual loss is applied to 3D-GRASE images, since noise is
removed, image details are mostly preserved and image contrast is more similar to that
of 3D-FSE images. The quantitative metrics used to evaluate the quality of the images
corroborated the higher image quality and similarity between the images enhanced by the
3D U-Net and 3D-FSE in the axial plane. However, the radiologist assessment indicated that
further developments need to be performed to improve the interface of the tissues in the
images from the networks to apply this work in the assessment of cartilage.
Finally, the contributions and conclusions of this thesis are discussed in Chapter 6.
Although further assessments need to be performed to adopt the technical developments of
this thesis in clinical settings, we have shown the benefits of 3D-GRASE with accelerated
reconstructed techniques over 3D-FSE and the promising capabilities of DL to enhance the
quality of 3D-GRASE images.

108 Summary

(SNR), and compared with similar 3D-FSE acquisitions. From this work it was concluded that
SORT Linear modulation encoding for T 2 -weighted images and SORT Radial modulation
encoding with M=0 for PD-weighted images obtain image quality comparable to 3D-FSE,
while reducing SAR by more than 40% and shortening acquisition time by 20%. On the
other hand, Chapter 4 investigates the effects of the different k-space trajectories proposed
in Chapter 3 together with two common k-space undersampling grids for CS combined with
PI (CSPI) in PD-weighted and T 2 -weighted 3D-GRASE acquisitions. CSPI requires an
incoherent undersampling, a variable density k-space grid and a fully sampled k-space center
in order to achieve an artefact-free reconstruction. Two undersampled grids proposed in
the literature for different sequences and applications fulfill these requirements: Variable
Density (VD) pseudo-random Gaussian grid and VD Poisson-disc grid. The incoherence
of the different combinations of k-space trajectories and undersampled k-space grids was
evaluated in simulation, phantom and in-vivo experiments, concluding that i) sampling
patterns combining a VD Poisson-disc k-space grid in both PD-weighted and T 2 -weighted
contrasts obtained the highest incoherence and ii) the trajectory has low influence on the
results.

Chapter 3 and Chapter 4 shown that the quality of 3D-GRASE images highly depends
on the trajectory applied during the acquisition. Moreover, the image contrast can slightly
change with respect to Fast Spin Echo (FSE) if the acquisition parameters are not carefully
chosen, due to the T -weighted contrast introduced by the Gradient Recalled Echos (GREs).
∗

In order to propose a solution to this problem, in Chapter 5 a Deep Learning (DL) method
that brings the appearance of 3D-GRASE closer to 3D-FSE images, removing artefacts and
achieving a more similar FSE image contrast, was investigated. Three different DL models
were developed based on a Three-dimensional (3D) U-Net in combination with three loss
functions previously proposed in the literature for regression problems: i) the voxel-wise
metric l 2 -norm, ii) Destructural Similarity Index (DSSIM) and iii) the perceptual loss. The
results from this work showed that the overall image quality in the axial plane is improved
when a 3D U-Net with a perceptual loss is applied to 3D-GRASE images, since noise is
removed, image details are mostly preserved and image contrast is more similar to that
of 3D-FSE images. The quantitative metrics used to evaluate the quality of the images
corroborated the higher image quality and similarity between the images enhanced by the
3D U-Net and 3D-FSE in the axial plane. However, the radiologist assessment indicated that
further developments need to be performed to improve the interface of the tissues in the
images from the networks to apply this work in the assessment of cartilage.
Finally, the contributions and conclusions of this thesis are discussed in Chapter 6.
Although further assessments need to be performed to adopt the technical developments of
this thesis in clinical settings, we have shown the benefits of 3D-GRASE with accelerated
reconstructed techniques over 3D-FSE and the promising capabilities of DL to enhance the
quality of 3D-GRASE images.