Red means it is quoted from Hornak; Green means it should be shown in class (e.g., picture, animation, table, etc.) The Basics of MRI by Hornak About the author How to get the book @ $20 per machine. My annotated version of his Contents (red highlight means included in course): Preface (Help) About the author 1. Introduction o o o o NMRI or MRI ? Opportunities in MRI Tomographic Imaging Microscopic property responsible for MRI 2. The Mathematics of NMR o o o o o o o o o Exponential Functions Trigonometric Functions Differentials and Integrals Vectors Matrices Coordinate Transformations Convolutions Imaginary Numbers The Fourier Transform 3. Spin Physics o o o o o o o o o o o o o o o o Spin Properties of Spin Nuclei with Spin Energy Levels NMR Transitions Energy Level Diagrams Continuous Wave NMR Experiment Boltzmann Statistics Spin Packets T1 Processes Precession T2 Processes Rotating Frame of Reference Pulsed Magnetic Fields Spin Relaxation Bloch Equations 4. NMR Spectroscopy o o o o o o Time Domain NMR Signal +/- Frequency Convention 90- FID Spin-Echo Inversion Recovery Chemical Shift 1 5. Fourier Transforms o o o o o o o o o Introduction The + and - Frequency Problem The Fourier Transform Phase Correction Fourier Pairs The Convolution Theorem The Digital FT Sampling Error The Two-Dimensional FT 6. Imaging Principles o o o o o Introduction Magnetic Field Gradient Frequency Encoding Back Projection Imaging Slice Selection 7. Fourier Transform Imaging Principles o o o o o Introduction Phase Encoding Gradient FT Tomographic Imaging Signal Processing Image Resolution 8. Basic Imaging Techniques o o o o o o o o Introduction Multislice imaging Oblique Imaging Spin-Echo Imaging Inversion Recovery Imaging Gradient Recalled Echo Imaging Image Contrast Signal Averaging 9. Imaging Hardware o o o o o o o Hardware Overview Magnet Gradient Coils RF Coils Qadrature Detector Safety Phantoms 10. Image Presentation o o o o Image Histogram Image Processing Imaging Coordinates Imaging Planes 11. Image Artifacts o o o o o o o o o Introduction RF Quadrature Bo Inhomogeneity Gradient RF Inhomogeneity Motion Flow Chemical Shift Partial Volume 2 o o Wrap Around Gibbs Ringing 12. Advanced Imaging Techniques o o o o o o o o o o o o o o o o o o Introduction Volume Imaging (3-D Imaging) Flow Imaging (MRI Angiography) Diffusion Imaging Fractional Nex & Echo Imaging Fast Spin-Echo Imaging Chemical Shift Imaging (Fat Suppression) Echo Planar Imaging (Functional MRI) Spatially Localized Spectroscopy Chemical Contrast Agents Magnetization Transfer Contrast Variable Bandwidth Imaging T1, T2, & Spin Density Images Tissue Classification Hyperpolarized Noble Gas Imaging Parallel Imaging Magnetic Resonance Elastography Electron Spin Resonance 13. Your MRI Exam o o o o o Introduction Screening The Imager The Exam Your Results 14. Clinical Images o o o o o o o o o Angiography Head & Neck Spine Extremities Glossary List of Symbols References Usage Statistics Software License 1. Ch. 1. Introduction 1.1. NMR versus MRI 1945: NMR discovered independently by Felix Bloch (Swiss) at Stanford and Edward Purcell (US) at Harvard (Nobel Prize in Physics for Bloch and Purcell, 1952). 1948: Felix Bloch publishes an equation explaining NMR—now called the “Bloch” equation. 1.2. History of MRI 1971: Raymond Damadian (US), proposed that cancerous tissue might have a different NMR signal from that of noncancerous tissue. 1972: Paul Lauterbur (US), Dept Chemistry, U Illinois proposed the first MRI method in a twopage article (Nobel Prize in Medicine, 2003). 1975: Richard Ernst (Swiss), Dept Chemistry, ET-H, Switzerland proposed the first MRI method employing a 2D Fourier transform, “Spin-Warp Imaging” (Nobel Prize in Chemistry, 1991). 3 1977: Peter Mansfield (UK), Dept Physics, U Nottingham, proposed the first “fast” MRI method, called “Echo Planar Imaging” (Nobel Prize in Medicine, 2003). 1.3. Important imaging techniques: 1.3.1. Spin-warp imaging = minutes 1.3.2. Echo-Planar Imaging (EPI) = fraction of a second. 1.4. Tomography = picture of a slice 1.4.1. CT (X-rays) 1.4.2. MRI (radio waves ~ 40 MHz, just to the left of the FM dial) 1.4.3. PET (gamma waves) 1.5. Digital images 1.5.1. voxel versus pixel 1.5.2. no color, just intensity 1.6. Meaning of intensities 1.6.1. CT: intensity = density of electrons 1.6.2. PET: intensity = density of radioactive material 1.6.3. MR: intensity = density of isolated protons 1.6.3.1. nuclei of hydrogen in water 1.6.3.2. nuclei of hydrogen in fat 1.7. MR machines 1.7.1. Uniform static magnetic field (the “static” field) 1.7.1.1. During all imaging, the static magnetic field is constant and uniform throughout the imaging volume. In fact one of the expensive aspects of MR machines is the achievement of a static field that is uniform within the imaging volume to 1 part per million (ppm). 1.7.1.2. We will henceforth assume that there is a static field, B, pointing in the z direction , throughout this course. 1.7.1.3. Super conducting: Never turned off (i.e., it is on 24 hours a day) 1.7.1.4. Resistive: Lower fields, usually large enough only for head images 1.7.2. Gradient: Controlled spatial variation in the static field 1.7.3. RF Pulse: Time varying magnetic field at about 40MHz produced by the “RF coil” or “transmission coil” 1.7.4. Receiver coil: Detects radiation from the protons. The intensity of this radiation is used to determine the density of the protons. 1.7.5. Show links to GE and Philips scanners 2. Ch. 2. The Mathematics of NMR 2.1. Look at the subsections. All should be familiar. Add the following: 2.1.1. Vectors 2.1.1.1. Know the definitions of the dot product: (1) involving their components, (2) involving the cosine of the angle between them. 2.1.1.2. Know the definitions of the cross product: (1) involving their components, (2) involving the sine of the angle between them. 2.1.2. Matrices 2.1.2.1. Know how to multiply a matrix times a vector 2.1.2.2. Know how to multiply two matrices, including the rule that makes multiplication legal 3. Ch. 5. Fourier Transforms 4 3.1. Students are responsible for knowing both the continuous and the discrete Fourier transform, convolution, and the FT of a square wave. 4. MR principles 4.1. Protons precess around the magnetic field direction (usu. z-direction) 4.2. Frequency of precession is proportional to field strength 4.3. When magnetic field from RF coil perpendicular to z rotates at precessional frequency, the protons absorb energy 4.4. When the rotating field is turned off, the energy is radiated at the same frequency 4.5. The receiving coil measures the strength of the radiation 4.6. With no gradient all protons absorb and radiate at the same frequency 4.7. With a gradient on, the frequencies vary spatially 4.8. The signal produced by the combination of frequencies can be Fourier transformed to show how the energy radiation is distributed spatially 4.9. The spatial distribution is the MR image 5. Ch. 3. Spin Physics 5.1. Spin: Each proton has a permanent “spin” of value 1/2. 5.2. Properties of Spin 5.2.1. Spin means that the particle has angular momentum and energy. It can change its energy level by absorbing a photon from radiation whose frequency B / 2 (Larmour equation), where gamma is the “gyromagnetic ratio”, where B is the strength of the magnetic field, usually given in Tesla. It ranges from 0.5 T to 3 T for clinical scanners. Vanderbilt is getting a 7 Tesla, whole-body scanner by the fall of 2005.or B , where “gamma-bar”, / 2 . For the proton, gamma = 267.8Mrad/s/T and gamma-bar equals 42.58MHz/Tesla. 5.3. Nuclei with Spin (omit) 5.4. Energy Levels 5.4.1. In a magnetic field, the proton can be in two energy states: Low when aligned with the field, High when opposed to it. 5.4.2. Transitions 5.4.2.1. When a photon is absorbed, the proton’s energy increases by E h , where h is Planck’s constant. When a photon is released, it decreases by the same amount. 5.5. Energy Level Diagrams (omit) 5.6. Continuous Wave NMR Experiment (omit) 5.7. Boltzmann Statistics 5.7.1. Protons in a magnetic field are continually absorbing and releasing energy, jumping between high and low states. 5.7.2. At equilibrium at absolute temperature T, the ratio: 5.7.2.1. N_opposed/N_aligned = exp -E / kT . 5.7.2.2. At room temperature, N_aligned is about 1/100,000 greater than N_opposed. 5.8. Spin Packets 5.8.1. The individual proton is governed by quantum mechanics, but fortunately, we can work with large numbers of protons at a time—large enough to permit us to use equations from classical mechanics. 5 5.8.2. 5.9. The protons produce their own magnetic fields, and their fields add or subtract depending on their directions. 5.8.3. In the presence of a magnetic field, they precess continuously about the field. 5.8.4. Magnetization 5.8.4.1. Because of random perturbations, their net field in the xy plane is zero, but the field in the z direction is proportional to N_aligned – N_opposed. 5.8.4.2. Because of the 1/100,000 larger value of N_aligned over N_opposed, there is a slight magnetization that adds to the impressed magnetic field. 5.8.4.3. The magnetization strength is labeled M0. 5.8.4.4. M0 is proportional to N_aligned – N_opposed. 5.8.4.5. So Mx = My= 0, and Mz = M0.= constant times (N_aligned – N_opposed). 5.8.4.6. The actual value of M is not important. All that typically matters are the ratios of its components to M0: Mx / M0, My / M0, Mz / M0, or the ratios of values of M0 at different static-field strengths or at different temperatures (see homework problem) or for different substances. Precession 5.9.1. (NOTE: I’ve changed the order from Hornak’s, which had T1 Processes before Precession. I did this after one of the students who had just heard my lecture on these two subjects in Hornak’s order ask questions. He and I realized that the order was confusing things. Also, I’ve added something about how the RF field puts the spins within the packets into phase, and I’m giving both a quantum-mechanical and a classical-mechanical description of what happens when a resonant RF field is applied.) 5.9.2. As pointed out above in “Spin Packets”, at equilibrium, Mx = My = 0. 5.9.3. Now suppose a second magnetic field (in addition to the static field B_0 in the z direction that we are assuming is always present throughout this course) of magnitude B1 is turned on perpendicular to the static field and rotating about the z-axis clockwise (as viewed from the top of the z-axis looking back toward the origin) at a frequency equal to the Larmour frequency. Such a field is applied for every MR image, and it is called the “transverse” field, because it is perpendicular to the static field, or the “Radio-Frequency field”, or, more commonly, the “RF field”, because it varies at a frequency equal to that of common radio waves. 5.9.4. Quantum mechanical description of the result: 5.9.4.1. This rotating field will cause the individual protons in the spin packet (a) to precess in phase with each other and (b) to cause the ratio N_opposed/N_aligned to increase, as protons aligned with the static field absorb energy from individual photons of the electromagnetic field produced by the rotating field and become opposed to the static field. 5.9.4.2. The absorption of energy by the protons is possible because the frequency of the rotating field is exactly correct to produce photons whose energy is exactly equal to the difference in energy between the aligned and opposed orientations. 5.9.4.3. Having these frequencies match is called “resonance”. 5.9.4.4. If the field is turned off when N_opposed/N_aligned = 1, then, Mz = 0, and, because the protons are precessing in phase, M rotates clockwise about the z axis in the xy plane. Since this rotation requires only about a hundredth of a second, the RF field that does it is called an “RF pulse”. 5.9.5. Classical mechanical description of the result: 6 5.9.5.1. This rotating field will cause the magnetization vector M to rotate about the magnetic field that is the total of the static field and the rotating field, and, as a result, its total motion will be to tip slowly and steadily away from the z axis while its xy component rotates about z at the Larmour frequency in the clockwise direction. The equation governing the rotation rate is dM t dt 5.9.5.2. 5.9.5.3. 5.9.5.4. M t B t (1) The steady tipping of M away from the z axis is possible because the rate of rotation of the field is exactly equal to the rate of rotation of M about the field. With any other rate, M will wobble around z but will remain near the z axis. Having these rates of rotation match is called “resonance”. If the field is turned off when the tip angle = 90 degrees, then Mz = 0, and M rotates clockwise about the z axis in the xy plane. As above, since this rotation requires only about a hundredth of a second, the RF field that does it is called an “RF pulse”. Show the second graphic in Ch 5 NMR Spectroscopy, The 90FID Sequence. 5.10. T1 Processes 5.10.1. The equilibrium magnetization vector (0, 0, M0) can be disturbed to a value (Mx, My, Mz) by radiating the protons with electromagnetic radiation at the Larmour frequency. 5.10.2. When the radiation is removed, the value M = (0, 0, M0) will be re-established over time. The evolution of the x and y components will be discussed below under T2 Processes. The z component’s evolution obeys the equation: M z M 0 1 e t / T 1 , where T1 is determined by the molecular environment of the spin packet. In NMR literature, T1 is called the “spin-lattice” relaxation time. Show the third graphic in Ch. 3. Spin Physics, T1 Processes. 5.10.3. While M0 varies somewhat with tissue type, T1 varies much more strongly. This variation is exploited (below) to make tissues have different intensities in MRI. 5.11. T2 Processes 5.11.1. After an RF pulse has been applied, the component of M that lies in the xy plane, sometimes called the “transverse” component, rotates at the Larmour frequency, but, also, its magnitude begins to attenuate. 5.11.2. The attenuation happens because the individual protons that comprise the spin packet begin to dephase. Show the third graphic in Ch 5 NMR Spectroscopy, The 90-FID Sequence. They dephase because they are exposed to slightly different magnetic fields. The difference happens for two reasons: 5.11.2.1. The fields caused by other protons in the molecules near a given proton within the spin packet vary temporally and spatially according to the protons’ orientations and how near or far they happen to be from the given proton. Hornak calls this effect the “pure” molecular effect. 5.11.2.2. The average static magnetic fields caused by electrons in the molecules near a given proton within the spin packet vary according to their orientations, their motions, and how near or far they happen to be from the given proton. This average field is major contributor to the magnetization of the object. As a result of this spatially varying magnetization, the local static magnetic field varies from one proton to another. Furthermore, the MR scanner itself suffers from 7 some spatial nonuniformity in its impressed static field. This effect is said to be due to static field “inhomogeneity”. (We will look at this effect in detail later.) 5.11.3. The resulting evolution of the magnitude of the component in the xy plane obeys the equation: M xy M xy 0 et / T 2 where (0) means the value of the component at t = 0 and T2 is determined by the molecular environment of the spin packet. Because of the universal use of the symbol T2, this phenomenon is often called “T2 decay”. 5.11.4. For all molecular environments, T1 > T2, so Mxy goes to zero faster than Mz goes to M0. 5.11.5. A common approximation (we’ll make it more exact later) is to write M xy M xy 0 et / T 2* , 1 T 2* 1 T 2 1 T 2inhom where T2 is due to the pure molecular effect and T2_inhom is due to static field inhomogeneity. In the NMR literature, T2 is called the “spin-spin” relaxation time. To emphasize that this dephasing is caused both by pure molecular effects and to static-field inhomogeneity, the decay is often called “T2-star decay”. 5.11.6. Like T1, T2 varies with tissue type, and it is another variation that is exploited (below) to make tissues have different intensities in MRI. 5.11.7. For B = 1.5 Tesla, typical values of T1 are given in a Table in Ch 4 NMR Spectroscopy, Problem 2. 5.12. Rotating Frame of Reference and Pulsed Magnetic Fields 5.12.1. It is mathematically convenient to define a coordinate system that is rotating relative to the stationary laboratory frame of reference clockwise at the Larmour frequency. In this rotating frame, Bz is the same as in the laboratory frame, but the xy component of B, B1, is stationary. We will assume, as most people do, that the orientation of the rotating frame is chosen so that Bx = B1 and By =0. In this frame, the only motion of M that takes place is as follows: 5.12.1.1. Before the RF pulse is applied, Mz evolves toward its equilibrium value, M0. 5.12.1.2. During the RF pulse, M rotates about the x axis. Show the first graphic in Ch 4 NMR Spectroscopy, The 90-FID Sequence. We will henceforth assume that the RF pulse is so short that the evolution of Mz toward M0 is negligible while an RF pulse is being applied. Thus, during the RF pulse, Mz decreases and My increases, while Mx remains equal to zero: M y M 0 sin B1t , M z M 0 cos B1t . 5.12.1.3. After the RF pulse, the evolution of Mz toward M0 starts again, M z M 0 1 e t / T 1 , and My begins to decay, M y M 0 et / T 2* , while Mx remains equal to zero. As mentioned above, the reason for this decay is the dephasing of the protons’ spins caused by spin-spin interaction. The transverse field, B1 will be left on for a time interval . If the interval is such that B1 is equal to pi/2, which is 90 degrees, then the magnetization of the spin packet is rotated into the xy plane, and the RF pulse is called a “90-degree 8 pulse”. There are 15-degree pulses, 20-degrees, and every other number up to 180. The most common are 90 and 180. 5.12.1.4. Henceforth we will assume that is not only much smaller than T1 but even much smaller than T2. 5.13. Spin Relaxation 5.13.1. Both T1 and T2 vary with temperature, with the strength of the static field B, and with tissue properties,. Ordinarily the temperature does not change appreciably in the human body, so it is not a factor, but it is important when ultrasonic heating is used to kill tumor tissue. By monitoring changes in the MR image it is possible to halt the heating before nondiseased tissue is harmed. Both T1 and T2 tend to decrease as the B field increases, but the field is so uniform across the scanner that it is not a factor within an image. It does change the contrast patterns from a low-field scanner to a high-field one. The major factor in imaging is the differences in T1 and T2 across tissue types. 5.14. Bloch Equations 5.14.1. The equation that Felix Bloch produced in 1948 describes the three phenomena of (1) the rotation of a spin packet around the magnetic field, (2) the evolution of Mz, and (3) the evolution of Mxy: (M x xˆ + M y yˆ ) (M 0 - M z )ˆz dM = g ( M ´ B) + , dt T2 T1 (2) where x̂ , ŷ , and ẑ are unit vectors in the x, y, and z directions. 5.14.2. When B varies with time, as it does, for example, when the RF pulse is active, there is no analytic solution to the Bloch equation. In that case, the Bloch equation must be solved numerically. Fortunately, for the purposes of this course, we will need to consider only special cases in which there is a solution. 5.14.3. The first special case occurs when M is at first directed along the z axis, and then an RF pulse is applied for a that is much shorter than T2. Since is much less than T2 and T1, we can neglect the second two terms of the Bloch equation. Thus, during an RF pulse: dM = g ( M ´ B) dt (3) Furthermore, it is not too difficult to show that in the rotating reference frame (see above) the equation becomes dM = g ( M ´ B1 ) . dt (4) 5.14.4. The second special case is the case in which B is constant with time and directed along z. There is, in that case, a simple analytic solution (see Problem Assignment 1). Also, there is a simpler form of the equation that holds in the rotating reference frame: (M x i + M y j) (M 0 - M z )k dM = + dt T2 T1 (5) 6. NMR Signals (i.e., without imaging) Ch. 4 NMR Spectroscopy 6.1. Getting a signal. We will go back to the laboratory reference frame and Eq. (2). Use the first and second graphic in Ch. 4 NMR Spectroscopy ,The Time Domain NMR Signal to illustrate how a signal is induced. (There is no slice selection here because we are not making an image!) 6.2. Because of the T2 decay, the signal gets smaller as it oscillates. Show the third graphic. Because the signal is produced by magnetic induction (from the protons to the receiver coil) and because there is no RF pulse driving the protons, this signal is called “Free Induction Decay”, or “FID”. 9 6.3. 6.4. 6.5. 6.6. As was mentioned earlier, the reason for the decay is two-fold: pure molecular effects and staticfield inhomogeneity. It is possible to “undo” that part of the decay that is due to the static-field inhomogeneity by means of a second RF pulse. Show the first four graphics of Ch 4, NMR Spectroscopy, The Spin-Echo Sequence. Because, in early experiments there was observed a second, weaker version of the FID signal after the second RF pulse, it was, and is still today, called a “Spin-Echo”. The time interval between the first RF pulse and the maximum of the echo, is called the “Echo time” TE. That time is exactly twice the interval between the first and second RF pulses. Many sequences, such as the Spin-Echo Sequence, or “SE” sequence, have been developed in NMR and MRI. To communicate a specific sequence, a “pulse sequence diagram” is used. Show the fifth graphic of Ch 4, NMR Spectroscopy, The Spin-Echo Sequence. Note that, in MRI, in order to get enough data to construct a complete image, it is necessary to repeat a sequence many times. The time interval between the first RF pulse of successive sequences is called the “repetition time”, or “TR”. Here are some sequences that you should be able to draw and for whose signal strengths, you should be able to give expressions: 6.5.1. Repeated 90-FID sequence. S k 1 e TR T 1 . 6.5.2. Repeated SE sequence. S k 1 e TR T 1 e TE / T 2 , where TE = the time interval between 6.5.3. the 90-degree pulse and the maximum of the echo. Repeated Inversion-Recovery (IR) sequence. S k 1 2e TI T1 e TR T1 , where TI is the “inversion-recovery time” time. Chemical Shift 6.6.1. There is one more field that affects the frequency of the protons’s precession: the electrons of the molecule that contains the H-atom. That field causes a shift in the Larmour frequency that is always the same for that molecule. Because it is related to the “chemical” nature of the molecule, it is called a “chemical shift”. Because the electrons reduce the field, the effective field has the form B B0 1 , where the value of sigma is on the order of a few parts per million. In the body, the two significant signal producers are water and fat. Their difference is about 3.5 parts per million (ppm). Since water is more important, this frequency shift is called the “fat shift”. It also has that name because the apparent position of fat is shifted in the image because of the frequency shift. 7. Ch. 6. Imaging Principles 7.1. Introduction (doesn’t say much) 7.2. Magnetic Field Gradient 7.2.1. A gradient is a spatial derivative of some quantity. In an MR scanner it is the derivative of the z-component of the magnetic field (when there is no RF pulse). 7.2.2. Electric current in special built-in coils, called “gradient coils”, that surround the imaging region can be turned on at varying strengths, positively or negatively so as to add a small magnetic field in the z direction that varies linearly across the field. The result is that the total static field B = B0 + Bgradient varies linearly by a few milliTesla across the imaging region. The gradients are designed such that at the center of the field B = B0, with B< B0 on one side and B> B0 on the other. 10 7.2.3. 7.3. There are three sets of gradient coils, named by the direction of the gradient that they produce: Gx, Gy, and Gz. They can be turned on singly or together to produce a gradient in B in any direction, but they are typically turned on singly. 7.2.4. While an image is acquired, gradients in all three directions are turned on and off at varying strengths. Because of the magnetic forces produced when a gradient coil is energized, there is a slight change in shape, which causes a loud “banging” sound, or a buzzing sound. This is the only sound a scanner makes, but it can be disconcerting to a person being scanned. There are no moving parts during the scan (except for the electrons moving to make the currents!) 7.2.5. Because any of the gradients can be turned on and off any order, there is nothing special about one orientation or the other. For example, a slice of the body that is perpendicular to the line from right to left can be acquired as easily as a slice that is perpendicular to a line running from feet to head. The former is called a “sagittal” view, the latter an “axial” view. The view perpendicular to the line running front to back is called the “coronal” view. The front is called “anterior”, the back is “posterior”, the direction toward the head is “superior”, toward the feet is “inferior”. Left to right is “lateral”. Frequency Encoding 7.3.1. When a gradient is applied, the spin packets at one position experience a different field from those at another. As a result, their Larmour frequencies vary across the field as well. For example, if the x-gradient is turned on with a strength Gx, then B x, y, z B0 xGx , v x, y, z B0 xGx , v0 xGx . At some values of x the density of spin packets will be higher than at other values, which means that M0 is different from one value of x to another. At the high-frequency end there may be, say, fewer packets, so M0 is smaller. At the center, x = 0, where the frequency is equal to 0 , there may perhaps be very few, even zero at some nearby point, and near the negative end of x, there may be a very large number of packets. If so, a plot of the density might look like this: The vertical axis is the number of spins per unit length, or it could be M0 per unit length, since they are proportional. The horizontal axis above is the x axis, but it could also represent frequency, because the frequency is proportional to x. Such a plot depicts “frequency encoding”, because the frequency of a spin reveals the position of the spin. It is this encoding that is the basis for all MR imaging. 7.3.2. As the magnetic vectors of these spin packets rotate, they radiate energy. The radiation will be at the frequency of rotation. In this case, there will be frequencies over the range, 11 0 FOV 2 Gx , 0 FOV 2 Gx , where FOV is the “Field of View”, which is the width of the imaging region. If the receiver coil is tuned so that it can pick up all the frequencies in this range, the signal that is received might look like this where the horizontal axis is time. In fact, two signals are picked up, one, sx(t), which is arises from Mx and sy(t), which arises from My. The signal shown above happens to be sy. The other signal sx might look like this: As described above, these two signals can be combined into one complex signal: s t sx t is y t As a result, the outputs of the two amplifiers that give these signals as outputs are called the “real channel” and the “imaginary channel”. The recording of these signals is called “readout”. The gradient that is turned on during readout, which in this example is Gx, is called the “readout gradient” or the “frequency-encoding” gradient. 7.3.3. The outputs of these two channels is digitized and recorded. A digital Fourier transform of s(t) is carried out by a computer (typically the one running the scanner itself): 12 S 0 F {s t } . S 0 is usually complex, but a plot of S 0 will look something like this: where the vertical axis is the amplitude of the Fourier transform. It shows a onedimensional image of the density of spins: an MR image! Field questions about HW Assignment 1 (~12 minutes) Reprise of what we’ve done so far: Driving force is to get a 3D map of the density of spin packets (worry about T1,T2 later) MR is good because it penetrates the body and is non-ionizing, but its wavelength is too long to permit our looking at transmitted or reflected waves Bloch’s equation shows (could demonstrate it in Matlab for example) that the resonance phenomena happens: Unless B1 is rotating at the Larmor frequency, M only wobbles: not much tilt. Maxwell’s equations (beyond our scope) show that if M rotates it will radiate. Need to excite only a region. Then we can infer that any radiation must come from that region. We can excite one slice at a time and then can frequency-encode in, say, the x direction. What’s left is the y direction That will be done using a new encoding method, called “phase encoding”, but before we look at that, we will look more carefully at how frequency encoding works. How frequency encoding works Sketch a pulse sequence that includes a phase-encoding y-gradient, but don’t explain how it works. Write out in integral form the relationship between the signal s(t) and the spin-density: s(t) = integral of spin density times exponential. Pull the common exponential with the central frequency outside the integral. Re-arrange so that the y and z integrals are part of the x-integrand and then define an “x-density”, which is a projection within the z-plane onto the x axis. Change variables so that the x-integral looks like an inverse Fourier transform of from x space into temporal space. Ask the class to accept on faith that it is easy by analog or digital transformations to move into the rotating frame, thereby removing the factor outside the integral. Point out that the Fourier transform of the signal is now seen to be equal to the x-density. Preview of phase endocing Now go back and explain that there will be a repetition of the sequence drawn on the board at the beginning in which the y-gradient will be changed. Give as an example the case in which it starts out large, becomes half size on the second repetition, goes to zero on the third, to negative half size on the fourth and to full negative on the fifth. 13 Suggest that we could acquire, during each readout, 5 (complex) numbers. Show a 5x5 array of dots representing the 25 values. Say that next time we will learn how to use a 2-D FT to change those 25 values into 25 intensities arrayed in x by y fashion. Correct the error of gamma that was originally in the notes, on the board, and in the homework. The change to these originals is that everywhere, except in the Bloch equations, , where 2 . I’ve already made that change in Homework Assignment 1 (by simply changing 2 to ). 7.4. Signal from a single spin packet. Because the signal is a combination of many spin packets with differing frequencies, the form of s(t) is complicated, but the form of the signal produced by a single spin packet is simple. Suppose the protons in an object inside the scanner are allowed to reach equilibrium, then a 90-degree pulse is applied, and then Gx is turned on. With t set to 0 at the moment that the pulse ends and Gx begins, the equations of motion for a spin packet at position x,y,z are M x M 0 sin 2 v0 xGx t e t T 2 , M y M 0 cos 2 v0 xGx t e t T 2 M z M 0 1 e t T 1 , The signal for this spin packet is s (t ) A sin 2 v0 xGx t i cos 2 v0 xGx t e t T 2 iA i sin 2 v 0 Ae xGx t cos 2 v0 xGx t e t T 2 i 2 v0 xGx t t T 2 e Notes about this formula. The i after the A is there because initially, the magnetic vector is lying in the y direction, instead of the x direction. We have absorbed i into A. The negative sign in the exponent means that the rotation is clockwise, instead of counter clockwise. If we ignore the T2 decay, the Fourier transform of this signal will be a single delta function at the frequency, v0 xGx . The T2 decay causes a spread in this delta function. The shape is the so-called “Lorentzian”. The kernel has the form c1 2 c2 2 . We will usually be able to ignore this spread when considering distortion problems. Ignoring the T2 decay, means that the signal for a spin packet is s t Ae i 2 v0 xGx t 8. FT Imaging Principles 8.1. Given that we have excited only one slice, it remains for us to produce a two-dimensional image of that slice. One dimension can be provided by frequency encoding. The second dimension is provided by means of “phase encoding”. 14 8.2. First, let’s look at a Fourier transform without considering MRI. 1D first: F f x , kx F kx f x e i 2 k x x dx f x F 1 F k x , x F k e i 2 k x x x dk And 2D: F f x, y , k x , k y F k x , k y f x, y e i 2 k x x k y y dxdy f x, y F 1 F k x , k y , x, y F k x , k y e 8.3. i 2 k x x k y y dk x dk y Suppose we excite a slice with a 90-degree pulse and then, before we turn on the readout gradient in the x direction, we turn on a gradient in the y direction. We leave it on for a time p and then turn it off. [Show a pulse sequence.] After that gradient is turned off, the signal produced by a spin packet at a position x,y,z will look like this Ae i 2 v0t yG y p The total signal will look like this: s t A x, y, za e i 2 v0t yGy p Aei 2 v0t x, y, za e dxdy i 2 yGy p dxdy As discussed earlier, we will assume that the high frequency oscillation is eliminated before processing by effectively acquiring the signal and transforming it to reference frame rotating at 0 . This transformation is accomplished by a device called a “quadrature detector”. Further information can be found in Hornak, Ch. 9, Imaging Hardware, Quadrature Detector. As a result, we can drop the exponential term in front of the integral signs to get, s t A x, y, za e i 2 yG y p dxdy We now turn on a frequency-encoding gradient in the x direction for a time f and let tr (the r is for “readout”) be the time elapsed since the beginning of that gradient (no gradient echo yet): s t f A x, y, za e i 2 yG y p xGx t f dxdy (6) for 0 t f f . While the frequency-encoding gradient is on, we read, or sample, the signal at each interval t : 15 sm s mt A x, y, za e i 2 yG y p xGx mt dxdy (7) We acquire N signals. Then, we apply another 90-degree RF pulse and repeat the whole process, but this time we change Gy. We will do this over and over, each time choosing Gy nGy : smn A x, y, za e If we define k x i 2 ynG y p xGx mt dxdy (8) Gxt f m Gx t and k y Gy p n Gy p , we have s k x , k y A x, y, za e i 2 k x x k y y dxdy (9) It can be seen that s k x , k y is the two-dimensional Fourier transform of the two-dimensional function of density, x, y, za . Because of the common use of the letter k in Fourier transform literature, the Fourier transform is said to existe in “k space”. If we take the inverse 2D Fourier transform of s k x , k y , we get A x, y, za . Take questions about HW 2. Draw pulse sequences on board for Problem 3 Put Eqs. (6)-(9) back on the board. (NOTE: It is easy to show that if a Fourier transform is taken of s k x , k y , instead of an inverse transform, then the resulting density is simply a 180 rotation of x, y, za about the z-axis.) A problem with the Fourier transform is that it cannot extend over infinity in the kx direction, because the image acquisition is only over a finite time. It is not practical to extend the acquisition over an arbitrarily large time interval because the signal dies out with T2 decay. After a while, the “signal” is nothing but noise. Also, we cannot measure the Fourier transform for negative values of kx, since time starts when we sample and is therefore nonnegative! We will see how to solve that problem with a so-called “gradient echo”. We can’t solve the problem of infinity, but we will later consider the effect of finite readout time on the image. 8.4. “Gradient echo” or “gradient-recalled echo”. Show or draw the third graphic (pulse sequence) in Ch. 8 Basic Imaging Techniques, Gradient Recalled Echo Imaging. Stipulate that the defocussing gradient extends over an interval that is equal to 12 f . It allows us to collect signal over twice the interval as without it. If there were no T2 decay, we would collect twice the signal, but because of T2 decay we get somewhat less than twice. 16 Change the origin of time tf to be the center of the readout interval, so that during readout, 12 f t f 12 f . Now look at the signal at the end of the defocusing gradient and the phase encoding gradient: s 12 f A x, y, za e Next look at the signal during readout: s t f A x, y , za e A x, y, za e i 2 yG y p xGx f / 2 dxdy i 2 yG y p xGx f / 2 xGx t f f / 2 i 2 yG y p xGx t f dxdy dxdy Be careful to explain that t f f / 2 t f f / 2 in the exponent in the first integral. Note that this looks just like Eq. (6), but the range of tf now includes negative values. If we repeat the sequence with different values of the phase-encoding gradient, as above, sample as above, and, as above, define k x Gxt f m Gx t and k y Gy p n Gy p , we again have s k x , k y A x, y, za e i 2 k x x k y y dxdy , but this time we have both positive and negative values of kx. Neither kx nor ky will extend to infinity, but they are centered around the origin. [Show a diagram of k space.] 8.5. Inhomogeneity. We have ignored inhomogeneity in the static field. Suppose there is a perturbation of the field B x, y, z . Then, it causes a continuous change in the phase from what we expect: s G y , t f A x, y, za e i 2 yG y p xGxt f B x , y , za t dxdy , where t is the elapsed time since the 90-degree pulse. The time elapsed between the 90-degree pulse and the echo is called the “echo time”, TE. Therefore, s Gy , t f A x, y, za e i 2 yGy p xGxt f B x , y , za t f TE dxdy At the center of the echo, where tf = 0, the signal is s G y , t f 0 A x, y, za e 17 i 2 yG y p B x , y , za TE dxdy (10) 8.6. Spin Echo. Show the first graphic (pulse sequence) in Ch. 8 Basic Imaging Techniques, Spin-Echo Imaging. Or draw one in which events take place in this order: 90-degree, phase encoding and xdefocusing, 180-degree, x-readout. Stipulate again that the defocusing gradient extends over an interval that is equal to 12 f but this time it is not inverted. Again, let the origin of time tf to be the center of the readout interval, so that 12 f t f 1 2 f . Call the time from 90 to 180 pulse TE/2. Point out that the gradient echo and the spin echo need not be coincident. Ask the class to comment on the restrictions necessary to get coincidence. ANS: The defocusing gradient can be anywhere before the readout gradient. The 180-degree pulse must be centered at the time that is exactly one-half of the time from the 90-degree pulse to the center of the readout gradient. Now look at the signal at the beginning of the 180-degree pulse. Ask the class what the effect on the maximum signal will be of a non-coincidence. Of a coincidence but at a longer TE. This is a good time to revisit what the source of the two types of T2 decay are. The signal looks like this at the instant before the 180-degree pulse: A x, y, za e i 2 yG y p xGx f / 2B x , y , za TE / 2 dxdy The 180-degree pulse now rotates the spins about the x axis 180 degrees which leaves the real component alone but negates the imaginary component, thus taking the complex conjugate gives the signal at the instant after the 180: A x, y, za e i 2 yG y p xGx f / 2B x , y , za TE / 2 dxdy Now look at the signal at the beginning of readout: s Gy , 12 f A x, y, za e A x, y, za e i 2 yG y p xGx f / 2B x , y , za TE / 2B x , y , za TE 2 f / 2 i 2 yG y p xGx f / 2B x , y , za f / 2 dxdy dxdy and at TE, the signal is, s Gy , t f 0 A x, y, za e i 2 yG y p dxdy This signal should be compared with that of Eq. (10). The missing B x, y, za means that the only dispersion in spins is that due to the phase-encoding gradient. It is for that reason that the 180-degree pulse causes a larger echo (larger than the mere gradient echo). The complex conjugate 18 means that the spins are rotated 180-degrees around the z axis relative to the situation without the 180-degree pulse. The signal during readout has the form: s Gy , t f A x, y, za e i 2 yGy p xGx B x , y , za t f dxdy We will examine the effect of the perturbed field on the reconstructed image in great detail soon, but for now, let’s assume that it is zero. So we have s Gy , t f A x, y, za e i 2 yG y p xGx t f dxdy We will also assume that the phase encoding gradients are reversed when the 180 is used, Gy nGy , so s Gy , t f A x, y, za e i 2 xGxt f yG y p s k x , k y A x, y, za e i 2 xk x yk y where k x dxdy, (11) dxdy, Gxt f m Gx t and k y Gy p n Gy p , as before, and G y is the value of the y gradient that would have been used if there were no 180 pulse (i.e., its sign is the opposite of the actual y gradient). As with the Gradient-Echo sequence, we have ignored inhomogeneity in the static field. Again, we suppose that there is a perturbation of the field B x, y, z . This time, however, the phase produced by the inhomogeneity during the time interval, TE/2, between the 90 and 180 pulses is exactly negated by the 180 pulse, at which time the sign of the exponent is changed. Then the same phase is produced at the same rate again, canceling more and more of the original. After the interval TE/2, at which time tf = 0, it will be completely and exactly cancelled. Thus, the signal looks like this for an arbitrary time tf during readout: s Gy , t f A x, y, za e i 2 xGxt f yG y p B x , y , za t f dxdy At the center of the echo, where tf = 0, the signal is s G y , t f 0 A x, y, za e i 2 yG y p dxdy , which is the same signal that we calculated for the Gradient-Echo sequence when there was no inhomogeneity! For comparison recall that with inhomogeneity, at the echo, the GE signal looked like this (Eq. (10) above): 19 s G y , t f 0 A x, y, za e 8.7. i 2 yG y p B x , y , za TE dxdy Field of View. All MRI machines perform an inverse Fourier transform to produce an image, and they perform such a transform digitally. The signal must therefore be sampled. The sampling interval t , the magnitude, G x of the frequency encoding gradient, the time p interval of the phase encoding gradient, and its magnitude, G y , determine the “Field of View” (FOV) of the image. To understand why, we must remember what happens with sampled Fourier transforms: The transforms are periodic. We will assume that that there are M samples collected during readout and N repetitions of the sequence in which the phase encoding gradient is changed, and we will assume that N is an even number. Then, m M / 2..M / 2 1, n N / 2..N / 2 1. It helps to define some convenient quantities: Y G y p ; 1 X Gx t , 1 x X M , y Y N ; m x / x, n y / y; Typically gradients, and time periods, are chosen so that X = Y, and the number of samples and number of phase encodings are equal, so that M = N, but we will keep them independent in our derivation. Returning to Eq. (11), and using these definitions, gives s kx , k y A x, y , z a e i 2 xk x yk y dxdy, excited region I m, n s m / X , n / Y B mx, ny e i 2 mm / M nn / N dmdn excited region We consider the integral to be approximated by a sum, as follows: I m, n C mx, ny ei 2 mm / M nn / N , m n where the sums extend over at least the excited region. Then, we can get an image by taking the inverse transform, The inverse transform looks like this: 20 i m, n D M 21 N 21 I m, n e i 2 mm / M nn / N , m M 2 n N 2 where i m, n mx, ny (and, although it is of little consequence, D C MN ) is the reconstructed image. [Point out that the upper case I is now suggestive of the Fourier Transform of the image, which is lower case]. We note that, because of the periodic nature of the discrete Fourier transform, i m, n D mx pX , ny qY p q i m M , n N , where the sums extend over at least the excited region. [Show a 2D repeating pattern.] If x, y is nonzero outside the Field of View— X 2 x X 2, there will be “Wrap Around”. Y 2 y Y 2 —then The top will show up on the bottom and vice versa and/or the left and right will overlap. The solution is to sample more often or to insure that the RF excitation does not extend beyond the FOV. Examples can be found on the web. Check out www.radiology.co.uk/srs-x/cases/084/d.htm. 8.8. Ch. 8, Basic Imaging Techniques, controlling contrast by adjusting TR and TE For Spin-Echo images S k 1 e TR T 1 e TE / T 2 and for Gradient-Echo images, S k 1 e TR T 1 e TE / T 2* . By adjusting TR and TE it is possible to emphasize the importance of T1, T2, or proton density. Because of T2* effects, GE images work well only for very short TE. Thus, they are used only for T1-weighted images or proton density. The limits below are not hard ones. Short TR (< 500 ms) Long TR (> 1500 ms) Short TE (< 25 ms) T1 weighted Proton-Density weighted Long TE (> 60 ms) T2 weighted Look at examples in Hornak: T1 weighted: Gradient Recalled Echo (TE=5 ms), theta = 90 degrees, TR = 50 ms PD weighted: Spin-Echo, TR = 2000, TE = 20 T2 weighted: Spin-Echo, TR = 2000, TE = 80 8.9. Ch. 8, Basic Imaging Techniques, Signal Averaging 21 To improve signal-to-noise levels, Number of excitations (Nex) is increased beyond 1. 8.10. Ch. 8, Basic Imaging Techniques, Multislice Imaging. I can’t do this better than Hornak. Herewith his own words, “An imaging sequence based on a 90-FID was introduced in Chapter 7. Based on this presentation, the time to acquire an image is equal to the product of the TR value and the number of phase encoding steps. If TR were one second and there were 256 phase encoding gradient steps the total imaging time required to produce the image would be 4 minutes and 16 seconds. If we wanted to take 20 images across a region of interest the imaging time would be approximately 1.5 hours. This will obviously not do if we are searching for pathology. Looking at the timing diagram for the imaging sequence with a one second TR it is clear that most of the sequence time is unused. This unused time could be made use of by exciting other slices in the object. The only restriction is that the excitation used for one slice must not affect those from another slice. This can be accomplished by applying one magnitude slice selection gradient and changing the RF frequency of the 90o pulses. Note that the three frequency bands from the pulses do not overlap. In this animation there are three RF pulses applied in the TR period. Each has a different center frequency 1, 2, and 3. As a consequence the pulses affect different slices in the imaged object. ” 8.11. Ch. 8, Basic Imaging Techniques, Oblique Imaging. Again Hornak does it very well: Slice Selection Gradient Gz = Gs Sin 60o Gy = -Gs Cos 60o Phase Encoding Gradient Gz = G Sin 30o Gy = G Cos 30o Frequency Encoding Gradient Gx = Gf 9. Ch 11, Image Artifacts 9.1. Bo inhomogeneity. We’ve already seen that it reduces the signal for GE sequences. Detailed analysis of that effect and the geometrical effect (and the Jacobian effect), which are present for both GE and SE, makes up the last seven weeks (half the semester) of the class. 9.2. RF inhomogeneity. See for example the paper by Likar Viergever Pernus (2001). See Figure 7 in that paper. 9.3. Motion 9.3.1. General (See Hornak’s graphics under Motion) 9.3.2. Cyclic: Gating can help (See Hornak’s graphics under Motion) 9.3.3. Flow: e.g. “Black blood” (See Hornak’s graphics under Flow) 22 9.4. Chemical Shift. The protons in fat are subject to a different effective Bo from those in water. Thus, their images “act” as if they are in the presence of an inhomogeneity B x, y, z cB0 , where c is approximately 3.5 x 10-6. The result is a geometric shift of the fat signal relative to its true location by x cB0 Gx , assuming that the readout direction is x (for non-EPI imaging). Here is the derivation that demonstrates it s Gy , t f fat A x, y, za e i 2 xGxt f yG y p cB0t f A x, y, za e i 2 A x, y, za e i 2 x1 x Gxt f yG y pt f xcB0 dxdy Gx Gxt f yG y pt f dxdy dxdy A x x1 , y, za e i 2 x1Gxt f yG y pt f A x x1 , y, za e i 2 x1Gxt f yG y pt f A x x1 , y, za e i 2 x1Gxt f yG y pt f dx1dy dx1dy dx1dy where x1 x cB0 Gx and x x1 cB0 Gx . Now, if we take an FT of the signal, we get i x1 , y, za fat C x, y, za C x1 cB0 / Gx , y, za Note that the “shift” is forward for c>0. We’ll do a similar, but somewhat more complicated, derivation when we study static-field inhomogeneity. 9.5. 9.6. Partial Volume. An effect from large slice thickness (See Hornak’s graphic under Partial Volume) that can cause blurry images and/or a superposition of anatomical features (that should be in two different slices). Gibbs Ringing. The result of the finite time of signal acquisition. Show Hornak’s graphic under this section. Since the acquisition is finite, the Fourier transform of the image is truncated. By FT theory, this means that the image is convolved with a sinc function. Looking back at Eq. (11), s Gy , t f A x, y, za e i 2 xGxt f yG y p s k x , k y A x, y, za e i 2 xk x yk y 23 dxdy, dxdy, we see that, in fact, s k x , k y is truncated in both the k x and ky directions. Draw pictures, a 2D example and a 1D example. 10. Ch 12, Advanced Imaging Techniques 10.1. Volume imaging (3D Imaging) 10.1.1. So far, we’ve seen slice-selection imaging, in which one slice is activated, signals are acquired from that slice and a 2D FT produces an image of that slice. By interleaving, it is possible to handle several slices at a time to produce a volume image, but only one slice is active at a time. Volume imaging makes it possible to produce a volume image by activating an entire volume. Phase encoding is performed in two directions (y and z) instead of just one with frequency encoding being performed also, as in slice-selection imaging. The entire volume contributes to each signal and the acquired signals are subjected to a 3D FT to produce the volume image. We can simply add now the z phase to the version of the expression for the signal given for the slice-selected image in Eq. (11) s Gz , , Gy , t f A x, y, z e s k x , k y , k z A x, y, z e i 2 xGxt f yG y py zGz pz i 2 xk x yk y zk z dxdydz , dxdydz , k x Gxt f m Gx t k y Gy py n Gy py k z Gz pz n G y pz 10.2. Fractional Nex & Echo Imaging Before starting this explanation show k-space and put in dots showing the points that need to be sampled before doing the inverse Fourier transform. Show which dots go with which time points for a standard GE sequence. We note that there is redundancy in the information that we acquire when we sample the signal over the entire readout interval. We get two numbers (real and imaginary) at every sampled time, which means that for SE or GE and for slice-selection or volume imaging, we have twice as many signal samples as we do image voxels. Mathematically, this can be stated in terms of the Fourier transform property that if is real, then s k x , k y s k x , k y * s k x , k y , k z s k x , k y , k z * 24 Thus, we can acquire half the samples and then calculate the other half. For example, in the slice-selection case, we can, for example, collect only signals for which k y is positive. Then, we can calculate the other values: s k x , k y s k x , k y , which in k-space means * that for each value below the horizontal axis we use the value of a point on the other side of the origin the same distance from the origin and take its complex conjugate. This technique is sometimes called, “Half-Fourier Reconstruction”. Hornak calls it “Fractional Nex Imaging”. (Nex means “number of excitations”). Alternatively, we can stop sampling at TE, to shorten readout time. Hornak calls this technique, “Fractional Echo Imaging”. The advantage is shorter imaging times. A drawback is that there is less signal, so signalto-noise is lower (by square-root of 2). Sometimes a few of the lines just below the axis or beyond TE are acquired and averaged in to get the low frequencies in the y, or x, directions, respectively. The result is that the SNR for lower (i.e., closer to zero) frequencies is higher, which is pleasing to the eye because more recognizable image content tends to reside in the lower frequencies. 10.3. Fast Spin-Echo Imaging After readout is completed, the spins are dephased and they are dephasing further, even after the readout gradient is turned off because of static-field inhomogeneity. Also, T2 dephasing is taking place. The former effect can be undone by another 180-degree pulse. Look at Hornak’s graphic under in Ch 12, Advanced Imaging Techniques, Fast Spin-Echo Imaging. HOWEVER, note that he has an incorrect sequence because he has a defocusing pulse before each readout pulse. In fact only one such pulse is needed (i.e, the one under the read phase encoding gradient is correct, but the rest of them should not be there). 10.4. Echo-Planar Imaging (Functional MRI) EPI is also called “Single-shot imaging”. Look at Hornak’s graphic in Ch 12, Advanced Imaging Techniques, Echo Planar Imaging (Functional MRI). Draw k-space on the board to the side of the screen, and trace it out while following the EPI pulse sequence. Discuss other ways to traverse the space: forward-backward, forward only, zig-zag, and spiral. 10.5. Variable-Bandwidth Imaging [This is a far simpler approach than that given in the (excellent) reference by Haacke, Brown, Thompson, and Venkatesan, Magnetic Resonance Imaging, Physical Principles and Sequence Design, Wiley-Liss, NY (1999).] Start with this argument in 1-D. The signal comes from the spins, whose number is proportional to x (ignoring the small changes in spin density), but the noise, which is unrelated to x , comes from everything in the room. As usual, we assume that the noise is uncorrelated, so its expectation value increases with the square root of the acquisition time. As usual, the signal is proportional to the acquisition time. Therefore, 25 Signal xN ex f ; Noise N ex f ; Signal x N ex f . Noise where we recall that f is the total sampling interval—the time over which the signal is read out. In 3D: Signal xyzN ex f N y N z ; Noise N ex f N y N z ; Signal xyz N ex N y N z f . Noise Note that we have used N x , N y here, instead of M , N , which were the symbols we used above. The value of N z is one for slice-selection sequences. Bandwidth. A commonly used term in dealing with SNR is the bandwidth. While the readout gradient is on, the Larmour frequencies vary across the field of view in the x direction. The difference between the highest frequency f max Gx X 2 and lowest frequency f min Gx X 2 , which equals BWread , or simply the “bandwidth”. Gx X called the “readout bandwidth”, Recalling previous the relation, X Gx t , we see that BWread 1 t , which equals the sampling frequency. This relationship could have been derived from the Nyquist theorem, which states that to capture frequencies (positive or negative) of magnitude up to and including f max , the sampling interval can be no greater than 1 2 f max . 1 Another commonly used term is the “Bandwidth per voxel” or “Bandwidth per pixel”, BW voxel BWread M . Using the relationships above we find that BW/voxel Gx x 1 f Therefore, we have that SNR xyz Nex BWread 26 MNN z where SNR means Signal-to-Noise Ratio. Thus, if the y-z quantities are fixed, then SNR x N ex BW voxel N ex x G x x N ex f Thus, it can be seen that the more spin packets ( x ) that contribute to the signal, and the longer the time of collection of signal from those packets ( N ex f ), the greater the SNR. On the other hand, increasing the bandwidth decreases SNR and increasing Gx decreases SNR. We remember from earlier that x B Gx is the erroneous shift in position in the x direction due to an error in the static field. Thus, we must sacrifice either geometric fidelity or SNR if we adjust Gx. To get both, we must do something else. 27