Mark S. Cohen

This brief set of notes describes some of the ways in which your choice of imaging parameters affects the overall quality of the images that you can obtain with an MRI system. As a radiologist, you will have several goals in mind in your selection of imaging parameters. On the one hand, you will certainly want images of the highest possible diagnostic utility. In general this refers to the best possible combination of contrast (based on the physiological parameters T1, T2 and proton density) combined with the best possible spatial resolution. Because most radiologists obtain their business largely through referrals, it is also desirable to obtain images of high aesthetic value, which are to the largest extent possible readable by the referring physician, who may not be well-trained in MRI. Finally, in order to make the MRI system financially practical, you will want to achieve the highest possible patient throughput consistent with these goals. The purpose of this document is to apprise you of some of the factors which are available to you to increase the quality of your images.

In general, it is safe to say that high image quality can often be achieved only at the expense of patient throughput, and one of the more difficult decisions you will have to address in day-to-day imaging is the selection of the most appropriate tradeoff between quality and throughput. In this paper are outlined some of the primary determinants of image quality with a brief explanation of the physical principles which yield the results of better (or worse) looking images. Where practical, a rational discussion will be attempted concerning the best possible compromises.

Among the major parameters affecting image quality are:

Signal to Noise Ratio

The signal to noise ratio (SNR) is a measure of the extent to which the MR image is contaminated by random noise. Noise in MR images shows up as an overall grainy appearance in the image: adjacent pixels on the video screen, even within the same tissue, have very different intensity. Noise of this sort ultimately compromises the conspicuity of fine details in the MR image and, in addition, makes the images aesthetically unattractive. Signal to Noise ratio is one of the more tangible parameters used to measure image quality. It is furthermore among the most limiting parameters of the MR experiment: magnetic resonance, by its very nature, tends to be SNR limited.

Contrast simply refers to the relative intensity ratio of two physiologically different tissues. A high contrast image is generally of higher diagnostic utility because it aids in the discrimination of tissues of various origins and in the distinction of diseased from normal tissue.

High contrast in the presence of a great deal of random noise is not useful. The random noise makes it difficult to determine whether intensity differences in the MR image are due to noise effects or to physiological differences. In producing high quality images it is our goal to achieve the highest possible contrast to noise ratio (CNR). Generally CNR is defined as:

(SA - SB)/N

where SA and SB are the signals from two tissues (A and B) and N is the background noise (or the standard deviation in pixel intensity).

The resolution of an image refers to the smallest possible feature that can be distinguished in an image. Of course we desire the finest possible resolution from our imagers. Resolution is ultimately determined by many factors.

Using a 256 x 256 matrix with a 50 cm field of view (a 50 cm object is divided into 256 different points in both the horizontal and vertical dimensions), the minimum feature size which we could distinguish would be:

50 cm/256 = 1.95 mm

If however, there is some distortion in which pixel a given tissue voxel is presented, then our actual resolution is reduced. Furthermore, if there is a large amount of noise in the image then it may be difficult to distinguish the boundaries of image features. In this way, a low signal to noise ratio compromises the image resolution. Slice thickness is a less visible dimension of spatial resolution, in the through plane dimension. Usually, MR images have anisotropic resolution: better resolution in-plane than through plane.

T2 decay for most tissues occurs with a time constant of 100 ms or less. In this time, the signal intensity is reduced to 37% of its starting value. After another T2 time constant (e.g. another 100 ms for tissue with a T2 of 100 ms) the signal has decreased to 14% of its starting value. The relative signal intensity of tissues with different T2 values (i.e. the contrast) increases as echo times are increased:

T2 Relaxation

At the same time that we are increasing T2 contrast, however, we are also losing total signal intensity and therefore losing out on signal to noise. When the MR signal has decreased in intensity to the point that it is at, or near the noise level of the system, our effective contrast is compromised as measured by the low contrast to noise ratio. Assuming that the TR values are long enough to minimize T1 contrast (see below), it is seldom necessary to extend the echo times beyond 70 to 90 msec to obtain very high T2 contrast. If there is significant contrast contamination due to T1 effects, then it may be necessary to extend the TE. This can only be done, however, at the expense of a loss in signal to noise ratio and image quality.

In summary, increasing the echo time (TE) improves T2 contrast. The maximum practical echo time is limited by the decrease in signal intensity, and consequently in signal to noise ratio, achieved with long echo times.

In order to achieve both high SNR and optimal T2 contrast it is important to choose a TR long enough that complete T1 relaxation has occurred between pulse repetitions. This ensures that the signal intensity is maximal and that there is very little intensity difference in the MR images due to variations in T1 between different tissues:

T1 Relaxation

The longest T1 relaxation times in the body are those of fluids, especially water and CSF. In general, T1 for CSF is about 1 second at a field strength of 0.5 Tesla and slightly longer at higher field strengths. According to the definition of T1, in one T1 time constant (e.g. 2-3 second for CSF) the longitudinal magnetization has returned to 63% of its final value. In spin echo images this also means that the signal intensity of a tissue has returned to 63% of its maximum value. As we wait for longer recovery period between pulses, more T1 relaxation occurs. After three T1 time constants (6 seconds for CSF) we have regained about 95% of our signal intensity.

Given that the noise in MR images is essentially independent of the imaging parameters TR and TE, the longer that we extend our pulse repetition time, TR, the better our signal to noise ratio.

In addition, when we wish to make T2-weighted images we have two primary goals, one is to have the maximum practical contrast between tissues with different T2 values, and the other is to minimize the contrast due to other parameters, especially T1. Therefore, in making a T2-weighted image, we wish to extend our TR times to the point that complete, or near complete, T1 relaxation has occurred for all of the tissues of interest. Usually this means waiting for approximately three times the longest T1 value of the tissues we are examining.

Increasing TR has the effect of prolonging the total scanning time, and it is this parameter which offers the most difficult choice for a compromise. T1 relaxation varies too, with magnetic field strength. Thus, for a 1 Tesla magnet, 95% relaxation may take 20-40% longer than for an 0.5 T instrument. Because of this, equivalent contrast weighting may require longer TR values.

Outside of the head, the tissues of interest in MR have somewhat shorter T1 times. In the body, T1 values are seldom above 700 ms. Thus, a TR of 2 seconds is sufficient for nearly complete T1 recovery, and high image quality.

In general, longer TR's achieve better T2-weighted and proton density-weighted images. They do so at the expense of increased scan time and consequently reduced throughput.

A T2-weighted image is by definition an image with high T2 contrast and minimal T1 contrast, a T1-weighted image has high T1 contrast and minimal T2 contrast. Finally, a proton density image is one with minimal contrast contributions from either T1 or T2 effects, and is thus assumed to show primarily differences in proton density. From these considerations the following chart can be established:

Setting up a chart of this kind begs the questions "What is a short TR? What is a long TE?" and so on. These are determined simply by the magnetic physics of the tissues being imaged. In general we can consider a TR to be long if there has passed sufficient time for 95% longitudinal recovery between each 90° pulse. With CSF having a T1 time constant of approximately 1 second (at 0.5 Tesla), T1 recovery requires approximately 3 seconds in this tissue. As TR is decreased, T1 contrast increases, and does so more dramatically as TR becomes close to zero. For our purposes, a TR of 0.5 second may be considered short.

Similarly, these variables can be considered for T2 behavior. The T2 time constants of most tissues of physiological interest range from about 25 to 500 msec. For high T2 contrast an echo time should be selected which is long enough to show contrast but not so long as to allow complete transverse demagnetization for the tissues we wish to differentiate (in the latter case, it may occur that two tissues of different T2 show up in the MR image as equivalently dark). In most cases a TE of 80 to 100 msec is appropriate to develop good T2 contrast, but the optimal TE value can be determined only by experimentation. The determination of a short TE is in some sense easy. A zero TE is without question optimal, but unfortunately impossible to achieve in actual imaging systems. For our purposes, a short TE will of necessity be defined as the shortest echo time achievable by the MR system. (A note in passing: in order to achieve very short echo times the MR manufacturer must make a variety of tradeoffs which will have side effects for other aspects of image quality. As these tend to be operating at fairly molecular levels of the equipment, they are beyond the scope of this overview paper, nevertheless, you should be aware that there is generally some deleterious effect for very short TE.)

Images with a TR of 2.5 seconds or more and a TE 30 msec are commonly called "Proton Density" images. It is worthwhile, however to note several point about these images. Firstly, let's examine their contrast behavior. If we are imaging a region having tissues with a T2 range from 30 to 100 msec (a reasonable spread of T2 values) the contrast ratio between the tissues (call them a and b) can be determined analytically by the following formula:

Which, if we plug in values for T2a of 100 msec and T2b of 30 msec, tells us that at a TE of 30 msec there is an intensity difference of 2 to 1 for these tissues. Thus, the proton density image can be seen clearly to have substantial T2 contrast, and is probably better understood to be a low—contrast T2-weighted image.

It is often thought that in a spin echo sequence, a second echo can be acquired for free - that is without detrimental effects on image quality. While this is not a terrible approximation to reality, it is also only an approximation. It is possible to design pulse sequences which produce better quality images by acquiring only a single echo, and in fact, such sequences are part of the standard software on many instruments. For example, in single echo sequences it is possible to obtain shorter echo times (and thus less T2-weighting) which results in better T1 imaging.

When acquiring T1-weighted images, you should always use a single-echo sequence. With the short TR values appropriate for T1 imaging a late (second) echo is of no value whatsoever. Thus it is desirable to optimize the single echo that you do obtain.

In long echo sequences (T2-weighting) the situation is more complex. The improvements in signal quality obtained by optimizing the pulse sequence for single echo acquisition are more subtle, but significant. In general the effect on the late echo, of acquiring a previous first echo, is an overall reduction in contrast. Since most physicians feel that the T2-weighted image is of the most clinical utility it is reasonable to do everything possible to improve that image. Thus it is at least worth some effort to experiment with the differences between single and double echo sequences using long TE's.

Due to the apparent effects of even echo rephasing (which make flowing blood appear brighter on second echoes than it does on first echoes), vascular detail is often better, especially in the brain, with first than with second echoes. When dark blood is desired single echo sequences are often better.

Since the strength of the MR signal is determined by the volume of tissue emitting that signal, thicker slices have better signal to noise ratio. They also, however, have reduced detail because of the effects of volume averaging. That is, the contributions of all of the elements throughout the thickness of the slice are superimposed, resulting in blurring of the final image. In this case, there is no simple formula for determining what slice thickness yields the best image quality. Often, 5 mm slices give excellent detail resolution while maintaining excellent image quality. It is probably a good idea to experiment with the effects of this parameter.

In the multi-slice imaging mode, if two adjacent slices receive 90° pulses at similar times then signal from one slice may contaminate the other. This results in a decrease in contrast and resolution. One way to combat this is to shuffle the order of slice excitations such that adjacent slices are not excited sequentially. Slice shuffling (performed automatically on most instruments) is a reasonably effective mechanism for reducing crosstalk if the TR is long enough. The improvement in image quality obtained is much less with short TR (i.e. T1-weighted) sequences.

Adding a gap of 40% or more between slices is extremely effective in improving image quality, but can only be done at the cost of incomplete volume coverage with each individual sequence

In the case of slice crosstalk, there appears to be a reasonable compromise: For long TR sequences (T2 and Proton density) use interlocking (assuming that contiguous slices are needed) to avoid crosstalk. The ability to acquire many slices in a long TR gives good volume coverage.

For short TR (T1-weighted) sequences the obtainable number of slices, and therefore the volume coverage is already reduced. In this case, multiple sequences will be needed to produce good volume coverage. Rather than acquiring two sets of contiguous slices, it is desirable to use two series each with a 100% gap, positioning the second multi-slice sequence to fill in the gaps in the first sequence. The total imaging time is unchanged over the acquisition of two un-gapped sequences, but the image quality is improved substantially.

As Field of View (FOV) is decreased, each tissue voxel is represented by more pixels in the final image and the signal from a given volume of tissue is spread across more pixels. The noise per pixel, however, remains constant. Thus SNR per pixel, and CNR, are compromised. Large zoom factors may ultimately have lower useable resolution (!) than images acquired without zoom. The bottom line: FOV decreases should be used conservatively. Typically a 24 to 30 cm FOV is desirable in head imaging, allowing the head of the typical patient to complete fill the image screen. In many cases, the overall appearance of a retrospectively magnified image will be superior to that of a zoomed image.

By averaging multiple images containing significant, but uncorrelated noise, it is possible to produce an improved SNR. In general, the SNR improves by a factor of Ã2 for each doubling of the number of averages, or NEX. Thus, a 2 NEX image yields a Ã2 improvement and the average of 4 images achieves a two-fold improvement in signal to noise as compared to an unaveraged image. Unfortunately, acquiring four NEX also increases the scan time by a factor of four. While averaging is sometimes absolutely necessary, it is a very costly, in terms of time, mechanism for improving image quality

Matrix Size
Moving from a 256 x 256 to a 128 x 128 (or 192 x 256 etc…) matrix has several effects. First of all the signal per tissue voxel is increase because the tissue voxels are larger. Secondly, the scan time is reduced by a factor of two. This two highly positive effects are offset by the fact that the minimum feature size is increased (that is, the resolution is decreased). The 128 x 128 matrix should be viewed as a practical alternative in certain difficult situations, such as pediatric cases, where the patient may not be able to sit still for long periods, or for zoom imaging, where it is desired to have an enlarged view of an organ or tissue without a severe loss of signal to noise. Furthermore, the anisotropic spatial resolution which results may yield unacceptably prominent "Gibbs" ringing artifacts at tissue boundaries

It is entirely possible to make a perfectly good image look terrible by setting the window and center parameters poorly before photographing it. Alternatively, careful attention to window setting may allow you to call the attention of a referring physician to a pathology which he or she might not otherwise have appreciated.

While this may seem a trivial point, it is one parameter in the production of high image quality which can be improved at no expense in patient throughput and can produce an enormous enhancement in perceived image quality. Make sure that your technologists understand thoroughly what you accept as proper windowing before photographing the images.

While your control over this important determinant of image quality is indirect at best, proper bedside manner can do much to relax a patient and to emphasize the importance of remaining perfectly still during imaging. The calmer the patient is, the more likely they are to remain motionless in the magnet. Anything that you can do to reduce their anxiety may be helpful.

While the service engineers working with your instrument are generally highly skilled in electronics and technical issues, they are not themselves clinicians and may be unable to judge good clinical image quality. In order for them to keep the machinery in peak condition it is helpful to keep a careful record of day-to-day performance and to thoroughly document any artifacts or service problems.

The day-to-day record usually involves the acquisition of a single image of a phantom at the same time each day that the equipment is used. This requires about ten minutes and provides an empirical documentation of any changes in basic performance.

Documenting the problems that occur includes noting the type of sequence being used when the failure occurred, the point during the operation when problems show up, and any error messages that the machine may print (although the machine should be keeping an internal record of these errors as well, it is helpful to the service personnel to correlate the messages with the actual failures.) In addition, it is often useful to save the raw data files should artifacts appear in the images. These contain an enormous amount of diagnostic information.

This final step in sending out the product of your imaging is often overlooked. Your camera must be calibrated carefully to ensure that the final prints look as much as possible like the images on the monitor screens. This calibration is an interactive process between the physician and the engineer installing the camera. Lastly, the choice of print film can make a surprising difference in the perceived quality of the MR images.

Some of the parameters mentioned here will have no effect on throughput, while others improve image quality only with an increase in total imaging time. These factors include:

Since time is money in MR imaging, I have ranked the remaining factors involved in quality improvement in increasing order of their cost:

A Final Note: Our understanding of the factors contributing to optimal clinical imaging sequences is constantly changing. While the protocols presented here are consistent with current understanding and technology they are not written in stone and should be appreciated as suggestions only.

rev 3/91