Our case study of magnetic resonance imaging (MRI) takes us from its scientific origins before World War II to the late 1980s. By 1988, MRI had achieved significant penetration of its primary market, clinical medicine. It had surpassed computed tomography (CT) scanners as the preferred diagnostic tool for a number of diseases, especially those affecting soft tissues of the head. Although the industry was in a shakeout phase, it appeared that a number of well-known manufacturers of medical instruments and, possibly, a handful of small entrants to the industry would survive and remain profitable. Worldwide sales in 1988 were approaching $1 billion annually, and about a third of the expected 4,000-unit market saturation level [31] had been sold (Lunzer, 1988). Cost of an installed unit began at $1 million, and a typical installation at a hospital or medical school could easily exceed $2 million.
Technologically, an MRI machine is basically a nuclear magnetic resonance (NMR) spectrometer with some important differences. MRI technology draws substantially from the science and engineering that created NMR. Yet, because the primary markets for NMR and MRI evolved in such a way that they overlap only slightly (analytical chemistry for the former, clinical medicine for the latter), the industrial histories are quite dissimilar. The MRI industry consists of a mix of companies well established in medical instrumentation (primarily X-ray machines and CT scanners) and a handful of new entrants established to take advantage of a perceived market opportunity. MRI machines sold in the late 1980s were a confluence of (1) a stream of science and engineering innovations related to NMR; (2) manufacturing, marketing, and engineering resources rooted in manufacturers of CT scanners[32] and X-ray machines; and (3) a small number of university-based researchers whose creative ideas and persistence brought the streams together in MRI.
A typical MRI machine for clinical application consists of the following technological elements: a large, powerful cylindrical magnet with a bore (aperture) large enough to enclose a human head or a reclining human body; a set of electrically conducting gradient coils that impose additional magnetic fields onto the object being imaged; a radiofrequency (RF) transmitter to excite nuclei in the object being imaged; an RF receiver to detect RF energy given off from nuclei as they return to their unexcited states; a computer to control the entire process, including the current fed to the gradient coils, the shape and timing of the RF excitation, and conversion of the pattern of RF energy received into a form that can be displayed as an image; and a display device such as a cathode-ray tube. Before describing each of these elements in detail, it is necessary to summarize the physical principles that underlie MRI.
Nuclear magnetic resonance was predicted by Wolfgang Pauli in 1924, observed in beams irradiated with RF energy in 1933 by Stern and Gerlach and in 1938 by I. I. Rabi, and observed in bulk materials by Felix Bloch and his team (William Hansen and Martin Packard) at Stanford and independently by Purcell and his team (Henry Torrey and Robert Pound) at Harvard in 1945. [33] Bloch and Purcell shared the Nobel Prize in physics in 1952 for their discovery. The resonance phenomenon depends on the fact that nuclei with a net charge have magnetic moments (spins). There are about 100 isotopes that have a magnetic moment, including 1H, 2H, 13C, 19F, and 31P. The fact that hydrogen is prevalent in animal tissue is highly significant for MRI applications in clinical medicine but of little relative importance for NMR applications in analytical chemistry. If nuclei possessing magnetic moments, such as 1H, are placed in a magnetic field, a small number of the nuclear magnetic moments will tend to line up parallel to the magnetic field, with slightly more along the field than opposite it. Like a gyroscope disturbed from a vertical orientation, a nucleus with a magnetic moment disturbed from its usual orientation by a pulse of RF energy will precess around the field lines of an imposed, static magnetic field with a particular frequency. The frequency of rotation, a function of the type of atom and the strength of the magnetic field, is called the Larmor frequency. [34]
When irradiated with RF energy at the Larmor frequency, atoms precessing in a magnetic field will absorb energy (resonate); when the RF energy is turned off, the nuclei will return to their normal state. The NMR signal is detected by RF pickup coils arranged to sense voltages induced by the transverse precessing magnetization. Irradiating an unknown substance in a magnetic field of known strength with a wide spectrum of RF energy will produce an output spectrum with peaks at the Larmor frequencies of each of the atoms with magnetic moments that comprise the sample. The NMR signal intensity is proportional to the density of atoms in the sample.
Subsequent to their 1945 studies, Bloch and his students found that the measured values of resonant frequencies varied slightly, depending on the chemical environment of the nuclei. In the case of 1H, for example, adjacent atoms shield the 1H nuclei to different degrees, causing the resonance frequencies to shift slightly. This phenomenon is known as the "chemical shift" effect. Although the effect initially was thought to be a confounding phenomenon, it was soon realized that by tabulating the chemical shifts for 1H and other NMR-sensitive isotopes in the presence of a wide variety of different atoms, the structures of complex molecules could be determined. This is the basis for high-resolution NMR spectroscopy.
Because slightly more than half the NMR-sensitive nuclei in a
sample align themselves along the direction of the magnetic field,
there is a net addition to the magnetization in that direction.
Changes in this additional magnetization occur when an RF field
is applied at right angles to the magnetic field. The processes
that occur when the RF field is removed are called relaxation.
Although not of primary significance for NMR spectroscopy, relaxation
processes proved crucial to NMR imaging (MRI). T1,
the spin-lattice relaxation time, is the time required for the
nuclei perturbed by the RF field to realign with the static magnetic
field. T2, the spin-spin relaxation time, is the time
required for the precessing nuclei to get out of phase with each
other as they transfer energy among one another and to adjacent,
nonenergized nuclei. Eventually, time-varying magnetization caused
by the precessing nuclei will decay to zero. Because MRI images
are most sensitive to 1H nuclei, they mostly reflect
the distribution of water in the sample. Since most MRI is done
on people, the H atoms in the water or fat of their body tissues
provide a highly sensitive focus for the technique. Different
tissues hold slightly different concentrations of water, but this
difference alone provides limited contrast in MRI images of animal
tissue. It turns out that both T2 and T1 relaxation
times differ far more widely across different tissues than does
spin density, so relaxation processes have been the focus of clinical
MRI applications. Both relaxation effects occur in all MRI imaging,
but because they can cancel each other, the researcher needs to
create conditions so that one or the other dominates.
An MRI machine consists of the magnet, gradient coils, RF transmitter and receiver, computer, software for performing a number of functions, and a display. The following is a more detailed discussion of each of these components sufficient to enable us to distinguish between intrinsic and supporting technologies. The former will then be the primary focus of our subsequent analysis of technological evolution of MRI and NSF's role in that evolution.
Magnets in commercially available MRI machines can be either superconducting, resistive, or permanent. More expensive machines with high field strengths (1.0-1.5 T) employ superconducting magnets; those with lower cost and field strength (0.15-0.50 T) tend to use resistive magnets. Resistive magnets are more unstable than superconducting or permanent magnets and are limited to a field strength of about 0.25 T. A few low-strength models use permanent magnets, but these are much heavier and limited in strength to about 0.2 T. The primary requirement of the magnet is that the field be as homogeneous as possible. Superconducting magnets offer higher field strength, and thus higher signal-to-noise ratios.
Superconducting magnets are often less expensive to purchase and maintain, especially since no other technology works above about 0.5 T. The development of efficient and cost-effective superconducting magnets has been an important advance in MRI technology. Recent refrigeration techniques reduce the need for liquid helium and thus widen the scope for use of superconducting magnets. Permanent magnets are expensive; resistive magnets require large amounts of power and closed-system cooling and are not currently competitive (information courtesy of William Edelstein, personal communication, October 29, 1996).
As of the early 1990s, superconducting magnets were used in 85% of the installations worldwide. The main suppliers of high-strength, precision magnets historically have been Oxford Magnet Technology, Ltd., a British firm that continues to dominate the market, and Intermagnetics General. By the mid-1980s, a few large MRI companies, such as GE, had moved from buying magnets from Oxford to manufacturing their own (Morris, 1992: 223-225). Oxford supplied magnets, refrigerators, and cryogenic equipment for most of the pioneers of MRI, either as part of a modified NMR spectrometer or separately.
These coils provide the spatially varying magnetic fields that are the key to identifying the locations of particular elements of a sample. Current machines have three coils, one each for the x, y, and z axes, with the z axis denoted by the direction of the static magnetic field. Early MRI experiments used only one coil that produced a static, linear gradient. Later developments led to multiple, time-varying gradients, so that drivers controlling the shape of the current supplying each coil became important. Requirements that the gradient fields be extremely stable and exhibit precise linearity presented formidable challenges for academic researchers during the experimental phase of MRI development and for industrial researchers and engineers during commercial development.
Early requirements for both NMR and MRI were for a highly stable source of RF energy, an engineering problem solved primarily by industry as demands for increased stability grew. Because demand was strong for stable sources of RF energy for experimental purposes in many fields and for applications such as radar and telecommunications, NMR and, especially, MRI researchers could draw on existing technology to build transmitters to the required specifications. For the same reasons, highly sensitive RF receivers were available or could be developed to meet researchers' and manufacturers' needs. Early in the development of MRI, however, it became evident that substantial advantages in scan time and image quality could be obtained if the RF energy was pulsed. Pulsed RF technology was already widely used in the NMR industry. As specialized pulse shape and timing requirements emerged for MRI applications, computer-controlled RF drivers had to be developed. Also, at high magnetic field strengths, special designs for the RF transmitter coil were required. The requirement for amplitude- and phase-modulated RF transmitter pulses was new and challenging.
MRI and related technologies such as CT scanners and NMR are completely dependent on computer technology. According to Blume (1992: 190), "The technology generally known as magnetic resonance imaging (MRI), like CT scanning, is inconceivable without modern computer technology." Morris (1992: 223) echoes this conclusion: "The importance of cheap computing power should not be underestimated: certainly it would not have been possible to have developed MRI techniques without it." Moreover, increased computer power and lower costs for equivalent performance probably accelerated adoption of MRI and made it accessible to a larger range of users. Although dependent on computers, MRI machines drew on existing computer technology at all stages of their development, rather than stimulating technological breakthroughs or significant developments in computer technology. [35]
MRI experiments and commercial machines employ, and always have employed, display technology already available on the market or developed for other purposes.
Unlike X-ray imaging, which produces an image directly on film, MRI produces an enormous amount of data that combine intensity and locational information for individual elements of a sample. There are several ways of accomplishing this combination, some of which derived from NMR research and others that are unique to MRI. NMR technology enables the spin density (quantity) of nuclei resonating at particular frequencies to be measured, leading to identification of the amount and identity of materials making up a sample. The incorporation of the magnetic field gradient in MRI enables locational information to be added to intensity data, but a two- or three-dimensional image of the sample must be reconstructed from the data.
In the projection method, a series of one-dimensional projections of position vs. intensity are produced by changing the orientation of a single gradient coil relative to the sample and making a large number of measurements, one set for each location. Using a principle first proposed by the mathematician Johann Radon in 1917, a complete, two-dimensional image of an object can be obtained from an infinite number of one-dimensional projections. A more manageable number of projections can produce a crude but identifiable image. [Interestingly, Radon's method was used before MRI in radio astronomy and to determine molecular structure from electron micrographs, but this experience remained isolated from initial application of Radon's technique to MRI (Morris, 1992: 222)).
In selective excitation methods, different techniques are used to excite nuclei selectively: point-by-point (sequential point methods), line-by-line (sequential line methods), plane-by-plane (planar imaging), and whole-volume (three-dimensional imaging) (U.S. Congress, OTA, 1984). Selective excitation methods typically require a combination of carefully designed and controlled RF pulses, varying gradient field strengths, and Fourier transform methods for converting the RF output information into forms that can be displayed on a screen. Fourier transform methods are preferable for most applications because they permit greatly reduced scan times (Morris, 1992: 223). Andrew noted in 1983 that at that time more than a dozen methods existed (Andrew, 1983). Since then, the pulsed Fourier transform method has become the dominant design and is used in most commercial MRI machines intended for medical applications. [36]
The two most common MRI pulse sequences are spin echo and gradient recalled echo. Within this broad classification, substantial manipulation of contrast can be achieved by changing echo time, repetition time, and RF pulse configuration, and by the inclusion of additional RF pulses for inversion or saturation of spins or the inclusion of gradient pulses. In this way, T1, T2, or diffusion-weighted images can be readily acquired. As Blume points out, the fact that images can be manipulated in numerous ways is a unique feature of MRI. Unlike CT scanners, "MRI does not image a conventionally unproblematic 'structure.' The image can be made to reflect one of a number of subtle, essentially chemical, properties of bodily tissue" (Blume, 1992: 220).
This section of our case study has described the physics underlying
MRI and the details of the several technologies that comprise
the types of commercial machines that have had, and are continuing
to have, significant influence on clinical diagnostics. Understanding
the historical development of MRI requires that a reader not only
grasp the underlying science and engineering, but also be familiar
with the historical development of NMR spectroscopy and CT scanners.
The contributions of MRI pioneers and the subsequent commercialization
of their ideas cannot be fully appreciated without knowing something
about how NMR spectroscopy and CT scanners came about.
We have already reviewed the physics underlying NMR and noted the scientific contributions of Pauli, Bloch, and Purcell. The history of NMR spectroscopy begins just before World War II in the Stanford physics department, which Felix Bloch had joined in 1934. William Hansen of the physics department was working on a "rhumbatron," a device for generating high-frequency radio waves whose frequency could be changed by physically changing its configuration. Russell Varian, already a prolific inventor and founder (with his brother Sigurd) of Varian Associates in Palo Alto, applied to Stanford for a Ph.D. and worked with Hansen in the physics lab. Russell Varian realized that Hansen's invention might be modified to generate the cloud-penetrating radio waves he and his brother knew could be used to make flying safer. The Varian brothers got an agreement with Stanford to use Hansen as a consultant and gain access to the physics department's equipment in return for half of the financial returns generated by what was to become the klystron, the RF source for U.S. and British radar during the war. By 1948, when Varian Associates was founded, close associations had been formed between the Varians and members of the Stanford physics department, particularly Hansen and Bloch.
Russell Varian was a research associate at Stanford when Bloch and Hansen conducted the experiments that would bring Bloch the Nobel Prize. Varian recognized that nuclear resonance would have application in analytical chemistry, and convinced Bloch and Hansen to file for a patent. Varian actually filed the patent, in return for an exclusive license from Bloch and Hansen to exploit their discovery. The Varian Instrument Division in 1948 was staffed by two former students of Bloch and Martin Packard, the third member of Bloch's team that conducted the 1946 nuclear resonance experiments. The company's major products were the klystron tube, NMR and EPR spectrometers, and electromagnets; eventually, Varian became the world's largest supplier of these devices.
The challenges involved in designing a good NMR spectrometer were to achieve high resolution, high sensitivity, spectral simplicity, and spectral stability. Resolution is a function of the stability and homogeneity of the static magnetic field; sensitivity is a function of the field strength of the magnet; spectral simplicity is achieved by obtaining the greatest possible chemical shifts, which is in turn a function of field strength; spectral stability is a function of the stability of the RF source used to excite the nuclei. These are, basically, engineering problems, so that "With such well-defined goals for an emerging technique of obvious analytical potential it was not surprising that commercial companies immediately took on the dominant role in the development of the technique" (Feeney, 1992: 207).
After a decade of development at Varian, progress in NMR spectroscopy was rapid, although somewhat overshadowed by the success of the klystron during much of this period. Varian introduced the first commercial NMR spectrometer in 1952 and, in 1961, the first commercial spectrometer using a superconducting magnet. By 1965, most universities were equipped with high-resolution NMR machines. A major breakthrough occurred in 1966, when Richard Ernst and Weston Anderson, both working at Varian, introduced the idea of using Fourier transforms to resonance spectroscopy (Ernst and Anderson, 1966). Rather than sweeping the sample with RF energy of continuously varying frequency, the sample is irradiated with a single pulse that excites all nuclei simultaneously. Using Fourier transform techniques, the resulting output from the sample is converted from time domain to frequency domain. The result is greatly increased sensitivity (signal-to-noise ratio) without changing any of the basic parameters of the NMR machine. For this and subsequent work on 2D and 3D resonance spectroscopy, Ernst was awarded the Nobel Prize in 1991. Fast Fourier transform techniques, developed at Bell Labs in 1965 by mathematicians Cooley and Tukey (Cooley and Tukey, 1965), greatly reduced the computing requirements of NMR. Bruker Instruments introduced the first commercial NMR spectrometer using Fourier transforms in 1969, and in the same year, computer-controlled pulse programmers were introduced. In the following year, Bruker introduced the first commercial Fourier transform (FT) spectrometer with a superconducting magnet. Thus by 1970, NMR spectroscopy was in widespread use in university and industry, applied to analytical chemistry.
In reflecting on the evolution of NMR, Feeney makes a useful distinction
between the relative roles played by what he calls "technology-driven"
and "curiosity-driven" developments. The basic discoveries
were made during 1945-53, then the innovation evolved for about
20 years through application of improved technologies, largely
developed in industry: superconducting magnets, computers, Fourier
transform methods, and pulse programmers (Feeney, 1992: 212).
After 1975, the origin of advances in NMR shifted back to university
laboratories, beginning with Ernst's development of multidimensional
techniques (Ernst had returned to his native Switzerland by this
time). At present, developments focus on multipulse and multidimensional
experiments; the technical history and current status can be found
in Ernst's Nobel lecture (Ernst, 1992).
Tomography has its origins among radiologists in the 1920s and 1930s who sought to deal with two shortcomings of X-ray images: there was no way to determine the depth of a pathology within the body, and there was no way to look "behind" bones. The problem was to introduce a third dimension into X-ray imaging. European radiologists developed tomography, in which both the X-ray source and film are moved successively in parallel but perpendicular to the object being scanned. Only one "slice" of the object is in focus at each position, so by building up a series of slices one could derive three-dimensional information.
In 1961, a neurologist at the University of California, W. H. Oldendorf, expressed dissatisfaction with traditional methods for producing images of the brain and proposed (in the IRE Transactions on Biomedical Electronics) a system that would produce a cross-sectional display of the head by recording differences in radiation transmission profiles and using them to calculate linear absorption coefficients. Oldendorf built a working model and applied for a patent in 1963, but his idea did not attract investors. The same year, a Tufts physicist, A. M. Cormack, performed experiments using these ideas and published the results in the Journal of Applied Physics, but this work, too, received little attention.
Godfrey Hounsfield, an electrical engineer, had been with Britain's Electrical & Musical Industries Ltd. (EMI) since 1951. After working in EMI's central research lab on computer memories, in the 1960s he turned to pattern recognition. This led him to consider the possibility of storing pattern information in a computer and subsequently eliminating redundant information before displaying it, rather than instantaneous imaging. He was aware of Radon's method of creating an image from an infinite number of one-dimensional projections, but thought in terms of building up a two-dimensional picture by accumulating data from a succession of one-dimensional projections taken from different angles. The two-dimensional picture would take the form of an n n matrix, with each cell of the matrix representing the density of one small element (pixel) of the picture. Hounsfield was already thinking of applying his idea to X-ray imaging.
By 1969, Hounsfield had convinced EMI to support construction of a prototype CT scanner and was operating it successfully. Internal funds were in short supply, so for subsequent models Hounsfield obtained partial support from Britain's Department of Health and Social Security (DHSS). In 1971, DHSS bought a prototype and underwrote the cost of constructing four more. By this time, the prototype scanner was being used in clinical trials in a London hospital and was attracting the interest of both radiologists and medical instrument makers. EMI introduced its first scanner in 1972, and in the same year EMI sold its first scanner in the United States. Hounsfield published a paper describing the instrument in the British Journal of Radiology in 1973. [39] By late 1975, 10 companies were offering scanners, but only EMI, Ohio-Nuclear, and Pfizer were actually delivering.
Georgetown University Medical School wanted its own CT scanner but thought the cost of about $350,000 was excessive, so a faculty member, R. S. Ledley, put together a group to build one that would accommodate an entire body. Ledley's design was basically EMI's but used a Fourier transform as the image reconstruction algorithm. The instrument was completed in 1973 and subsequently patented. Ledley set up his own company, which was quickly bought out by Pfizer. Not surprisingly, EMI later sued for patent infringement.
Meanwhile, in 1975, EMI introduced a whole-body scanner. Ohio-Nuclear, affiliated with the Cleveland Clinic, also entered the market. In 1977, EMI continued to dominate the world market with a 60% share, and by 1978, 15 companies had entered the market. These companies were of two types: established X-ray equipment manufacturers who felt threatened by the new technology and firms seeking to enter a new market in the hope of quick profits. General Electric entered the field in 1978 and quickly achieved market dominance through a combination of market recognition carried over from its X-ray machines, fan beam technology[40], and, as a large R&D-intensive company, the ability to focus technical and financial resources quickly. EMI sued Technicare of Ohio-Nuclear, Pfizer, and GE for patent infringement in 1977 and 1978, but EMI suffered major financial losses in 1978 and 1979. The company was acquired by Thorn Electrical Industries in 1979. Thorn-EMI's market share fell to 37% in 1980. At this time Thorn-EMI tried to sell its scanner interests to GE, but GE was advised by the U.S. Justice Department that this might violate antitrust law, so GE acquired all of EMI's non-U.S. operations. Part of GE's deal was to resolve the patent suit with EMI.
By 1977, a combination of factors had led to declining sales and
profits, signaling the beginning of an initial shakeout of the
industry. In EMI's case, the company was unable to continue to
invest the huge sums in R&D required to stay at the forefront
of technology. Profits shrank as the market tightened in response
to government (both U.S. and U.K.) efforts to control health care
costs. For the most part, it was the X-ray instrument suppliers
who had the resources to remain viable. This situation was already
being reflected in the market positions of leading firms in 1979:
Sales ($M) |
|
| General Electric | 100 |
| Siemens | 50 |
| Johnson & Johnson (Ohio-Nuclear) | 50 |
| Picker | 30 |
| Pfizer | 30 |
| EMI | 15 |
| Elscint | 14 |
| Philips | 10 |
| Toshiba | 5 |
| Hitachi | 5 |
Thus the CT scanner represents the contribution largely of industry
researchers and engineers, drawing on some long-standing ideas
and existing technology. Success in the marketplace was due primarily
to the name recognition in hospitals and medical schools of established
X-ray instrument manufacturers who had the financial resources
to withstand the shakeout. The relevance of CT to MRI is that
CT introduced the concept of computerized imaging rather than
direct imaging, as in photography. Also, the profits that accrued
to CT manufacturers such as General Electric provided the resources
to mount an intensive development of MRI, once its promise became
evident (SRI interview with Walter Robb, March 19, 1996). The
evolution of MRI resembles a combination of the intellectual history
of NMR and the industrial experience of CT scanners, but with
some interesting twists.