Several imaging modalities have been developed with the aim to acquire signal from different physical properties of the scanned anatomy, providing functional (e.g. PET, SPECT) or structural information (e.g. MRI, CT). Their signals complement each other, therefore it is becoming common for radiotherapy treatment to involve a combination of modalities. This provides more information about the anatomy at the cost of more work spent on time-consuming manual data processing.

The primary goal of radiotherapy is to design a treatment plan to cure, shrink, or stop the progression of cancer [115]. This requires a good anatomical understanding of the tumor and the surrounding organs at risk. The two most common imaging modalities are MRI and CT both with its unique strengths and limitations. MRI excels in providing high contrast for soft tissue, whereas CT relies on ionizing radiation for diagnostics and calculating the effects of the delivered dose during radiotherapy.

Computed Tomography

Computed Tomography (CT) is an imaging technique based on ionizing radiation, using X-rays to produce cross-sectional images of the scanned anatomy. The scans are often used for diagnosis and the monitoring of treatments, and since it provides information about the relative electron density of the scanned anatomy, the scans can be used to deliver the most effective radiotherapy treatment by calculating the dose exposure of the tumor and surrounding organs [83].

The signal acquisition in CT involves the detection and conversion of X-rays transmitted through the patient into electric signals, which are then reconstructed into images of the scanned anatomy. The pixel intensities in CT images are called CT-numbers, which use the HU [57]. The CT-number is calculated from the attenuation of the scanned anatomy, scaled to the attenuation coefficients in air and water, which makes the pixel values of CT absolute, and directly comparable between scans. As the HU is related to the relative electron density of the anatomy, which is related to the absorbed dose in the tissue, the CT scan can be used to calculate the dose distribution of a planned treatment [68].

The CT scans for our projects were acquired at the University Hospital of Umeå, Sweden with a Philips Brilliance Big Bore (Philips Medical Systems, Cleveland, OH, USA).

MRI

MRI is a medical imaging modality that uses a strong static magnetic field in combination with time varying magnetic fields, to create a non-ionizing method to quantify the amount and molecular composition of hydrogen atoms in the scanned area by detecting subtle changes in their magnetism [105, 41].

Due to its ability to image soft tissue with high contrast, MRI holds an important position for diagnostic and radiotherapy purposes. Additionally, the modality has the possibility to arbitrarily position the imaging planes, unlike CT. The non-ionizing property is another advantage of MRI, valuable in pediatrics or in sensitive scenarios where CT cannot be used.

Magnetic Resonance Physics

Placing a subject in a strong, static magnetic field B_0 induces a net macroscopic nuclear magnetization inside the scanned anatomy. This is caused by the abundantly present hydrogen atom inside the human anatomy, more specifically the hydrogen nucleus (a single proton) and its property called spin, which has its foundations in quantum mechanics.

Spin

The proton has an electric charge, and due to the configuration of the quarks that make it up, it also has a non-zero angular momentum. This property, the spin entails that the proton cannot perfectly align with the static magnetic field B_0, hence it experiences a torque expressed as:
\tau = \mu \times B_0,
where \mu is the magnetic moment of the proton. To explain the magnetization of the individual protons would involve quantum mechanical formulations, however the main concepts around the image acquisition process can be described using a classical mechanical-physics formulation. For this, assuming that there is no interaction between the individual protons, we can define the sum of the magnetic moments for a system of protons as:
M = \mu_0 + \mu_1 + \cdot = \sum_i \mu_i .
The net magnetization M describes a large sample of protons (in the magnitude of 10^{23}).

Despite the quantum mechanical nature of the interactions between a magnetic field and the protons, the dynamics of the net magnetization M in the magnetic field B are described well on a macroscopic scale by the Bloch equation:
\frac{d M}{d t} = \gamma M \times B.

Due to the non-zero angular momentum, the net magnetization cannot align with the main field. The magnetization in the z direction (of the magnetic field B_0) is expressed by the Bloch equation as
M_z(t) = M_z(0),
whereas the torque \tau will cause M_{xy} (the component of M in the transverse plane), to precess around the direction of B_0 as
M_{xy}(t) = M_{xy}(0)e^{-i \omega_0 t},
where \gamma = 2.7\times10^8~rad~s^{-1}~T^{-1} is the gyromagnetic ratio, and \omega_0 is the frequency of the precessing of the proton, called the Larmor frequency, defined as \omega_0 = \gamma B_0.

The dynamics of M during equilibrium are visualized in Figure 1.


Field strength
Figure 1. Larmor precession of the macroscopic magnetization oriented parallel to B_0. Due to the spin of the individual protons, the magnetization cannot align with a static magnetic field, causing it to precess with the Larmor frequency, which scales with B_0.

RF Pulse

A RF transmitter coil can be used to create an RF field B_1 with a specified frequency \omega_{RF}. Since this is a time-varying field and can be turned on and off, it is usually called an RF pulse. For the case of \omega_{RF} = \omega_0, the RF pulse is at resonance, causing the macroscopic magnetization to tilt into the direction orthogonal of B_0—driven by subtle quantum mechanical interactions. This means the decrease of the magnetization in the direction of B_0 (M_z), and the increase of the transverse magnetization (M_{xy}). The angle between M and B_0 is called a flip angle and is defined as:
\alpha = \gamma B_1 t_p,
where t_p is the time of the pulse. Commonly used RF pulses are usually denoted by their flip angle, such as a 90^\circ pulse (where \alpha = 90) or 180^\circ pulse (\alpha = 180). The time of the pulse and the strength of B_1 is carefully tailored to give the desired flip angle for each scenario.

Apart from tilting the direction of the magnetization, the RF pulse also leads to the phase coherence between the microscopic magnetizations, which leads to an increased net magnetization. Without phase coherence, the different magnetizations would be out of phase, and therefore cancel each other out. Instead, after a 90^\circ pulse, M_z = 0, the net magnetization precesses around B_0 and the individual magnetizations are in phase coherence. At this point, the RF pulse is turned off. This oscillating magnetic field can be used to induce voltage in a receiver coil, which is the basis of the MRI signal.

The phase coherent transverse magnetization will produce the MRI signal, not permanently however. The magnetizations continuously return to their natural state, which means they begin to dephase, and realign with B_0. The effects of the RF pulse on the magnetization can be seen on Figure 2.


Field strength
Toggle RF Pulse:
Figure 2. The effect of an RF pulse on the macroscopic magnetization. The magnetization is rotated towards the transverse plane of B_0 as the RF pulse is applied, and as the pulse turns off, the precession axis of the proton re-aligns with B_0 due to relaxation.

Relaxation times

The effects of an RF pulse that are utilized during signal acquisition are better described by the magnetizations of systems of protons, as visualized in Figure 3. In this scenario, we first decompose M into M_z and M_{xy}, a component in the direction of B_0 and in the transversal plane, respectively. The M_{xy} component can be further decomposed into the individual magnetizations, showing how they lose phase coherence after the RF pulse.

Before the RF pulse is applied, the net magnetization aligns with B_0, which means M_z = M. As the RF pulse is applied, the magnetization is rotated towards the transverse plane of B_0, and for the case of the 90^\circ pulse, the RF pulse is turned off as the magnetizations align with the transverse plane (M_z = 0) with phase coherence.

After the pulse is applied, M_{xy} (the magnetization in the transverse plane) will induce voltage in a receiver coil, and the signal starts decaying as the magnetizations dephase and start to realign with B_0. The differences in relaxation times between tissues will lead to different signal intensities, contributing to the high contrast between tissue types that makes MRI so useful.


RF Pulse:
Figure 3. An alternative visualization of the effect of the RF pulse, now using 10 individual magnetizations. The phase of the magnetizations align in the transverse plane and M_z decreases to zero as the RF pulse is turned on, and dephasing begins as the pulse is turned off. The left and right graphs show M_z and M_{xy} plotted against time, respectively. They reflect the state of the visualized magnetizations of systems of protons.

The relaxation entails two mechanisms: the realigning with B_0, and losing the phase coherence. The previous Bloch equations do not account for these re-alignments, however additional relaxation terms an be added to both.

As the magnetization begins to align with B_0, M_z increases. This is known as spin-lattice, or T_1 relaxation. This is caused by an increasing number of protons returning to their state parallel with B_0, and the energy released by the protons is distributed in the body as heat, increasing the body temperature by a very small amount. The changing of M_z defined in the previous respective Bloch equation can be expanded with a term that accounts for the observed T_1 relaxation as:
M_z(t) = M_z(0) (1 - e^{-t / T_1}),
where t is the time since the RF pulse has been turned off, and T_1 is a property of the scanned tissue.

The other mechanism that takes place is that the magnetizations begin to lose their phase coherence, returning to their equilibrium state—, also known as spin-spin or T_2 relaxation. A term which accounts for T_2 relaxation can be added to the changing of M_{xy} in the previous respective Bloch equation, making it:
M_{xy}(t) = M_{xy}(0) e^{-i \omega_0 t} e^{-t / T_2},
where T_2 is the property of the scanned tissue. One cause of T_2 relaxation actually comes from the energy release during T_1 relaxation. This means that the processes causing T_1 relaxation always results in T_2 relaxation. However, T_2 relaxation can also be caused by other processes, such as local field disturbances, or a pair of spins exchanging their longitudinal angular momentum. This means that in general, T_2 \leq T_1.


RF Pulse:
T1 Relaxation Time:
T2 Relaxation Time:
Figure 4. The T_1 and T_2 relaxation times of the scanned tissue affects how fast the spins realign with B_0 and how fast they dephase with each other, respectively. The left and right graphs show M_z and M_{xy} plotted against time, respectively. They reflect the state of the visualized magnetizations of systems of protons.

The T_1 and T_2 relaxation times depend on the scanned tissue, spanning between a few milliseconds (e.g. bone), to a few seconds (e.g. pure water). In general liquids that are made up of small, rapidly moving molecules (e.g. water, Cerebrospinal fluid) have high T_1 and T_2 times, which means they relax slowly. Whereas dense solids (e.g. tendons, proteins) have very short T_2 values, meaning the spins dephase quickly. For most tissues, the T_1 relaxation times are around 10\times larger than T_2.

The measured magnetization therefore provides useful knowledge about the scanned anatomy as it depends on the T_1 and T_2 relaxation times, as visualized in Figure 4.

Echoes for imaging

The relaxation times are not measured directly, but inferred from the acquired signal. As the magnetization dephases, it can be refocused by gradients and RF pulses to collect the signal. This refocusing effect that causes the signal to get stronger again is called an echo, with the two most common types called gradient and spin echoes.

Gradient echo

After the RF pulse is turned off, a special coil is used to create a gradient speeding up the dephasing of the protons. Afterwards, the gradient is reversed, creating a negative gradient which causes the previously dephasing protons to rephase.

Accelerating the dephasing of the protons with a gradient is required for spatial encoding, however refocusing becomes more difficult, as more time passes between the end of the RF pulse and the echo. The relaxation caused by the positive gradient is actually a composite effect of T_2, the inhomogeneities of the B_0 field and the susceptibility of the tissue, which is called effective T_2 relaxation time, denoted as T^*_2. The stages of a gradient echo sequence are shown in Figure 5.


RF Pulse:
Positive gradient:
Negative gradient:
Figure 5. Gradient echo sequence. The left and right graphs show M_z and M_{xy} plotted against time, respectively. They reflect the state of the visualized magnetizations of systems of protons. As the RF pulse is turned off, the spins begin to dephase. This can be accelerated by turning on the negative gradient. After a while, if the gradient is reversed, by turning on the positive gradient, the dephased spins begin to rephase. As time passes, the T_2 relaxation, independent of the gradients, will add to the dephasing of the spins, making their refocusing less likely.

Spin echoes

Spin echoes begin with a 90^\circ RF pulse, after which the spins start to naturally dephase. Applying an 180^\circ RF pulse will flip all the spins, so that the spins dephasing clockwise will now continue dephasing counter-clockwise with the same speed. After the same length of time as between the 90^\circ and 180^\circ pulse, the spins will refocus, forming a spin echo. The stages of a spin echo sequence are shown in Figure 6.


RF Pulse:

Figure 6. Spin echo sequence. The left and right graphs show M_z and M_{xy} plotted against time, respectively. They reflect the state of the visualized magnetizations of systems of protons. As the RF pulse is turned off, the spins begin to dephase either clockwise (shown in blue) or counter-clockwise (shown in red). Applying the 180^\circ pulse, their direction changes, as they eventually refocus and start dephasing again.

Spatial encoding

In a constant magnetic field B_0 we can use the above described techniques to obtain information about the protons inside the system, in the form of magnetization in the transverse plane detected by receiver coils. However, this setup cannot assign a spatial location of the protons based on their magnetizations detected by the coils. For MRI to be useful, the acquired information about the protons should be separated based on their spatial origin.

The acquired MRI data is collected in the form of the k-space, which gives spatial information about the scanned anatomy, encoded into the signal using both its phase and the frequency.

This is achieved by three gradient coils G_x, G_y, G_z which slightly increase the field strength depending on the spatial position of the proton along the x, y and z axes, respectively. This will also change their precessing frequency.

Slice selection

As described before, the RF pulse needs to be exactly the same as the Larmor frequency \omega_0 to excite the protons. Therefore, if the gradients increase the Larmor frequency of the protons as we move along some selected direction (in our example this is z), then an RF pulse with a specific frequency will only excite the protons at a specific location of z. With adjusting the frequency of the pulse, we adjust which z location to sample from, which defines the selected 2D slice. Additionally, adjusting the bandwidth of the RF pulse can also be used to select the thickness of the slice excited.

This is the first layer of spatial encoding in image acquisition. As the RF pulse is applied we also apply the slice selection gradient which will limit the received information by the coils to a single slice, as seen in Figure 7.


Gradient

Pulse Frequency

Figure 7. The selected slice is visualized by the red overlay while the currently acquired line is visualized by the dark red segment. Without the gradient G_x, there is no spatial encoding, the acquired data is from the entire system. As we increase the gradient, the frequency of the RF pulse defines the acquired slice.

Determining the orientation of the gradients will also determine the orientation of the slices. Gradients that change along the medial-lateral direction will create slices in the coronal plane, gradients that change along the anterior-posterior direction create slices in the coronal plane, while gradients that change along the superior-inferior direction will create axial image slices. The directional terms are visualized in Figure 8.

Figure 8. Directional terms of the human anatomy.

Phase encoding

With phase encoding we can deduce where the data came from along one direction in the excited 2D slice, using a gradient orthogonal to the slicing direction. In this example the phase encoding (PE) direction is selected to be y. This gradient will encode the position into the signal based on the phase of the spins, according to their position in y. After the RF pulse is applied, their phases are in coherence, however by introducing G_y, the protons will dephase from this original direction according to their location in y.

Let’s assume a uniform distribution of protons. We see that by increasing G_y we get a periodicity of protons being in phase amplifying the signal, and canceling each other out resulting in a net magnetization of 0. This of course changes when the distribution of protons is not uniform, however covering a wide range of G_y strengths, the behavior of the signal describes the protons along y canceling each other out or amplifying each other, which encodes the information about the protons based on the spatial frequency of the system. A simple demonstration of the spatial frequencies can be seen in Figure 9.




Pulse Frequency
Figure 9.A visual example of phase encoding. If all five protons are present, going through a range of frequencies shows a periodicity of the phases amplifying and cancelling each other. This periodicity changes if one of the protons is removed. Covering a wide range of frequencies we can infer the existence and abundance of protons.

The number of different values of G_y defines the resolution along that direction, i.e. for a resolution of 256, the same number of different values of G_y need to be applied. The time between acquiring two lines of data is called a repetition time (T_R) and it’s one of the scanner settings.

Frequency encoding

The same phase encoding technique can be applied in the remaining frequency encoding (FE) dimension, in our case x. However instead of going through the range of frequencies one-by-one, we can apply the gradient G_x and measure different dephased signal while collecting the signal from the echo created for the current G_y. This leads to a significant speedup, as this encoding can be performed during a single echo, essentially for free.

The time between the RF pulse and the created echo is called an echo time (T_E) and it depends on the used echo sequence, and it is one of the scanner settings defining the characteristics of the output image.

Acquired data

The acquired 3D data is collected in the frequency domain, the k-space. To transform the frequency signal into spatial signal the Fourier transform can be applied on the data. While the middle of the image space contains the signal from the middle of the scanned spatial domain (the middle of the anatomy), in contrast, the middle of the k-space shows the middle of the frequency domain, so the signal from the lowest frequency content, visualized in Figure 10.


Acceleration Factor

Figure 10. Collecting a smaller part of the center of the k-space leads to a more blurry signal, as higher frequency information is missing.

Data collection is faster in the FE direction than in the PE direction. Hence, it leads to a more significant speedup to collect less of the k-space in the PE direction, as seen in Figure 11.


Acceleration Factor (PE)

Acceleration Factor (FE)

Acquisition time:

Figure 11. Collecting fewer lines in FE direction gives almost no speedup but loss of quality. It is more sensible to downsample the PE direction instead which as the quality decreases also leads to significant speedup.

Image acquisition

Gradient and spin echoes can be used to collect information about the magnetization of the system, while gradient fields can be used to connect the measured magnetization to spatial locations. The variety of available sequences have a wide range of applications: while very short sequences are beneficial in rapidly moving anatomies, longer sequences generally provide an improved image quality with less noise.

Contrasts

Two essential settings that define the reconstruced image are connected to how the echoes are created. For both gradient echo and spin echo procedures, the echoes are repeated multiple times to increase the signal, and to enable the encoding of the positions. The time difference between the RF pulses is the repetition time (T_R), while the time between turning off the RF pulse and achieving the echo is the echo time (T_E).

The selected T_E and T_R will affect the signal according to the T_1 and T_2 relaxation times of the protons. The signal equation of a spin echo sequence is the following:
s = PD \cdot \left( 1 - e^{-\frac{T_R}{T_1}} \right) \cdot e^{-\frac{T_E}{T_2}},
where PD is proton density, another tissue-dependent property. The selection of T_E and T_R will define the contrast of the acquired image.

The two settings add weights on how much the three quantitative information will be present in the output image according to the signal equation. A short T_R will decrease the weight of the T_1 relaxation times, while a short T_E will decrease the weight of T_2. If both are decreased, the output signal is called PD-weighted. The effects of the two settings are visualized in Figure 12.


Echo Time [TE]

Repetition Time [TR]
Figure 12. Pelvic MRI, with customizable T_E and T_R times to show how these settings change the characteristics of the image based on the T_1 and T_2 relaxation times. The left and right images show T2– and T1-weighted contrasts, respectively.

Imaging trade-off

The data acquisition of MRI is slow, compared to e.g. CT. Although techniques are often used to speed up the acquisition time (e.g. parallel imaging), as discussed in the previous parts, both the individual slices and the phase encoded 1D lines of the k-space need to be collected in succession. This means the acquisition time increases linearly with the number of slices, and the number of pixels in the PE direction of the slice.

In MR imaging, a constant challenge lies in the balance between three variables: SNR, resolution and imaging time. While advancements in imaging techniques and hardware components may improve all three aspects of the system, prioritizing one of these qualities inherently results in a trade-off with the other two. Therefore, the choice of the sequence must be adapted to improve the quality that is most important for the specific case.

Hardware

The MRI equipment detects very subtle changes in magnetism. Therefore, despite the advanced engineering solutions and constantly improved imaging techniques that make the images more detailed, the many components of the scanner also lead to drawbacks and increase the possibility of artefacts.

The scans for the presented studies have been collected on a 3T Signa PET/MR scanner (GE Healthcare, Chicago, Illinois, United States) at the University Hospital of Umeå, Sweden.

In general, the MR scanner has three main components; the magnets for generating B_0, the gradients for spatial encoding, and the RF system for generating and detecting the signal. These components can be seen in Figure 13.

Figure 13. Schematic drawing of the components of a PET/MRI system—which is what we used throughout our projects.

Magnets

The previously described steps for data acquisition all heavily rely on the assumption that B_0 is constant inside the system, most commonly 3T or 1.5T. However, since the main field is induced through a coil (cooled down to its superconductive temperature), the slightest imperfections can cause inhomogeneities in the B_0 field.

Gradients

Three independent coils are responsible for the gradient fields (G_x, G_y and G_z) that are used for localization and for generating echoes. It could be assumed from the previous description that the gradients are turned on and off instantaneously, however, in reality, they have rise and fall times, for the coils to reach the specified strength. During these times the gradients do not increase and decrease linearly, which can introduce distortions in the reconstructed image. However, these distortions are relatively easy to correct for. More significantly, Eddy currents are produced in various places when the gradients are switched on and off, which disturb the spatial encoding of the signal. These currents can lead to artefacts, particularly during fast imaging techniques, and they are hard to predict and correct for.

RF system

The RF coils are placed inside the gradient coils, and for a PET/MR system (which is what we used throughout our projects) the PET ring is also integrated into this coil system. As with the main magnetic fields and the gradient coils, small imperfections in the RF system can lead to unexpected signal. These artefacts in the reconstructed signal, known as bias fields, might be caused by non-homogeneous transmitter coils or spatial variations in the sensitivity of the receiver coils.

Patient safety

The imaging technique is non-ionizing, however, there are other safety concerns related to the MRI system. Some of the main concerns and how they are mitigated are listed below:

  • Ferromagnetic objects can act as projectiles near the scanner. Hence, there are safety precautions around the MRI scanner, with the scanner being in a separated room where no magnetic objects are allowed.
  • Metallic implants such as pacemakers may be affected by the MRI scanner. They might move due to the strong magnetic field, overheat due to the RF, or stop working. More recent implants are designed so they can be scanned without any complications.
  • The force produced by the currents in the gradient coils cause an excessive loudness during the imaging process. This might cause discomfort during the scan. This issue might be mitigated by earplugs or earmuffs.
  • Peripheral Nerve Stimulation (PNS) caused by the rapid switching of the gradient fields can lead to muscle contractions. Although not dangerous, PNS can be very uncomfortable and can make it difficult for the patient to lay still.
  • A long-term concern about MRI is the accumulation of the contrast agents, e.g. gadolinium [72], often used to enhance the visibility of certain structures or tissues within the body. However this is only an indirect effect of the imaging technique and it can also be avoided by the use of multiple contrasts replacing the contrast agent, if the concerns increase.
  • The exposure to RF pulses leads to localized heating. This exposure can be measured and therefore limited by the Specific Absorption Rate (SAR), however, these are constantly monitored during the scan, and the exposure cannot exceed the safety limit.

In a controlled environment and with careful consideration of a safety guidance practice, the MRI system is considered very safe [105].

Artefacts

The MRI system is sensitive, small imperfections in some components can lead to significant artefacts in the reconstructed image. The image artefacts in MRI can generally be categorized into three main classes: hardware-related, sequence-related or patient-related [105, 173].

Hardware

Noise

Electronic noise, the motion of molecules in the patient, or the inevitable resistance of the coils, can all introduce a random Gaussian noise in the collected data. When the k-space signal is converted to an image, the distribution is skewed into a Rician distribution. This unstructured noise lowers the SNR of the image, and it can be mitigated by increasing the acquisition time or by using denoising algorithms during signal reconstruction.

Bias

Tissue-specific radio frequency penetrations and inhomogeneities of the B_0 and B_1 fields can affect the strength of the signal depending on its spatial origin. This breaks the assumption that a homogeneous tissue will result in a homogeneous signal. Instead, the bias introduces a low-frequency multiplicative noise over the image which leads to some parts of the image being brighter or darker independent of the scanned anatomy. Vendors frequently have built-in solutions that correct for the non-uniform receiver-coils, and retrospective bias field correction algorithms are also commonly used [151], most commonly the N4ITK algorithm [160], with many tuneable parameters.

Sequence

Gibbs ringing

As detailed above, collecting a larger k-space improves the reconstructions of high-frequency changes in the image. Consequently, when the collected k-space is not large enough, sharp changes in the image intensities might not be recovered well enough. The undersampled k-space introduces a ringing effect, known as a Gibbs artefact or Gibbs ringing [44]. The artefact stems from the Fourier transform that reconstructs the image from the collected k-space data. It is a common artefact in undersampled images, and it can be mitigated by increasing the resolution of the signal.

Aliasing

If the collected k-space does not physically cover the scanned anatomy, i.e. the anatomy extends beyond the collected data in the PE direction, the aliasing artefact will introduce an image wrap round. This artefact can easily be avoided by sampling lines in the k-space more densely for collecting more data.

Patient

Motion

Some imaging sequences can take more time than others, with a common scan spanning typically between 10 and 15 minutes. However during one exam, multiple sequences might be performed. Patients might find it difficult to remain still for the time of the scan for multitudes of reasons, e.g. pain, claustrophobia, discomfort, or simply because they are not used to it. Although the phase encoding stage of the acquisition is relatively fast, movement between acquiring these lines of data can introduce ripple effects in the final image. Techniques to speed up the acquisition time can make the scan more comfortable for the patient.

Metallic implants

Non-ferromagnetic medical implants such as dental fillings are allowed inside the scanner, however they can also lead to issues during imaging. They generally have higher susceptibilities than the surrounding tissue, therefore they cause magnetic field inhomogeneities leading to a loss of signal around the implants.

Current radiotherapy workflow

In Sweden, more than 60,000 patients are diagnosed with cancer every year [36]. The two most common cancer types, similarly to other Nordic countries, are breast cancer and prostate cancer [9]. The most common cancer treatment is surgery, however, more than half of all diagnosed patients also receive radiotherapy as a part of their treatment. Radiotherapy can be used as a standalone treatment, but it can also be combined with other cancer treatments, such as surgery or chemotherapy.

The goal of radiotherapy is to kill, shrink, or stop the growth of malignant tumors, and to reduce the risk of cancer spreading to other parts of the body. This is achieved by delivering a high dose of radiation to the tumor, while minimizing the dose to the surrounding healthy tissue. The radiation is most often delivered through a photon beam, but other particles (e.g. protons) or heavy ions can also be used [97].

With the quality of imaging systems constantly improving, the role of medical imaging in radiation therapy treatment planning is steadily increasing. The radiotherapy treatment planning—at the University Hospital of Umeå, Sweden—commonly starts with an MRI and a CT scan of the patient. The images of the two modalities are derived from two separate physical properties, and they also serve different purposes: organ delineation and dose planning.

Organ delineation

The MRI scan offers higher soft tissue contrast and is therefore used to delineate the anatomy of the patient, identify the tumor and surrounding healthy tissue. Often multiple scans of different contrasts—such as T_1-weighted, and T_2-weighted—are used to improve the delineation of the anatomy.

High accuracy delineations are necessary for high accuracy treatment planning, to calculate how much dose will be delivered to the tumor and to the surrounding healthy tissue. An essential element of organ delineation is identifying possible surrounding healthy organs sensitive to radiation, so the treatment can be planned to avoid them [67]. The identified tumor is first defined as the GTV. This delineation is expanded by a carefully defined certain margin to account for possible microscopic spread of malignant cells. This expanded margin defines the CTV. This volume is further expanded to account for all other uncertainties in the treatment (e.g. positioning, registration errors) defining the PTV that will be treated.

Dose planning

The treatment plan aims to deliver the prescribed dose to the PTV while delivering as little dose as possible to the surrounding healthy tissue. Using the MRI for delineations together with the CT for dose calculations, the MRI needs to be registered to the CT image frame of reference since the two scans are performed in different scanners, and the movements of the patient can introduce non-rigid deformations between the two scanned volumes [161]. To ensure that the organ delineations and the dose distributions are aligned, the two scans are registered retrospectively. The combination of modalities greatly increases the effectiveness of the treatments [125, 132], however possible errors in the registrations introduce additional uncertainties in the precision of the treatment plan [131].

MR-only workflow

The superior soft tissue contrast of MRI and the possible uncertainties introduced by the registration process of MRI and CT have led to the development of MRI-only workflows [68]. Here, the CT scan is replaced by an MRI scan that is used both for delineation and dose calculation [114]. One popular approach for this is to convert the MRI scan into an sCT scan, providing synthetic HU values, which is then used to calculate the delivered dose.

Although the modalities contain different physical information about the anatomy, ML has been very successful at converting MRI images into sCT images [66]. This solution eliminates the need for a CT scan for radiotherapy treatment planning, the manual work of registering multiple modalities and therefore speeds up the workflow.

Thesis content:

References