01Executive summary
Three milestones that made the interior of the body, and then the sequence of its instructions, into computable data.
Godfrey Hounsfield’s scanner produced the first clinical cross-sectional images in 1971, turning imaging into a reconstruction problem rather than a shadow-casting one. The first human magnetic resonance images followed in 1977. A reference human genome sequence was published in 2003, after which sequencing cost fell faster than semiconductor cost over comparable periods. All three are cases where the instrument produces measurements and the image or result is computed.
02Computed tomography: the image is a solution, not a picture
A conventional radiograph is a shadow. Everything along each ray is superimposed, so depth information is lost and a dense structure hides what lies behind it. That limitation is inherent to casting a shadow, not to the equipment.
Tomography measures attenuation along many paths through the body at many angles, then solves for the distribution of attenuation coefficients that would produce those measurements. The mathematics for this had existed since Radon’s work early in the twentieth century. What was missing was computation cheap enough to perform the reconstruction, which is why the milestone falls in 1971 and not decades earlier.
The instrument does not measure the quantity of interest. It measures line integrals through it, and the quantity is recovered by solving an inverse problem. Once that structure is recognised, the engineering priorities change: sampling geometry, conditioning, noise propagation and regularisation become the design variables, and detector improvement matters mainly through its effect on the inversion. The same structure appears in seismic imaging, electrical impedance tomography, radio astronomy and industrial process tomography.
03Magnetic resonance: making frequency mean position
Nuclear magnetic resonance was a laboratory spectroscopy technique for decades before it became imaging. In a strong static magnetic field, hydrogen nuclei precess at a frequency proportional to field strength. Apply a radio pulse at that frequency and they absorb energy; they then re-emit as they relax, and the relaxation rates depend on the local molecular environment — which is why soft tissues differ so clearly.
The step to imaging was spatial encoding. If a controlled gradient is superimposed so field strength varies with position, then resonant frequency varies with position too. Frequency becomes a proxy for location, and by applying gradients in different directions and combining the acquired signals, a spatial distribution can be reconstructed. Paul Lauterbur introduced gradient-based spatial encoding and Peter Mansfield developed methods for rapid acquisition and reconstruction; they shared the Nobel Prize in Physiology or Medicine in 2003.
Priority in this field is disputed. Raymond Damadian demonstrated that relaxation times differ between tissue types and patented an approach to detecting that difference, and objected publicly and at length to his exclusion from the 2003 prize. The defensible position is that Damadian established a key physical basis for tissue contrast, and that the spatial encoding which turned resonance into imaging came from Lauterbur and Mansfield. Both contributions were necessary, and this series does not attempt to adjudicate the award.
Field homogeneity
Position is inferred from frequency, so any unintended field variation is a geometric distortion. Superconducting magnets are shimmed to extraordinary uniformity and the room is screened, because the physics of the measurement makes uniformity an accuracy requirement.
Gradient switching
Rapid gradient switching produces the characteristic acoustic noise, induces eddy currents in surrounding conductors, and can stimulate peripheral nerves. Slew rate is limited by patient safety, not by amplifier capability.
Radiofrequency heating
Energy deposited by the excitation pulses heats tissue. Specific absorption rate is monitored and limited, which bounds how rapidly some sequences can run.
The projectile hazard
The static field is always on and is strong enough to accelerate ferromagnetic objects violently. Access control to the magnet room is a hard safety requirement, not a procedure.
04Sequencing: a cost curve steeper than semiconductors
The Human Genome Project ran for over a decade at very large cost and produced a reference sequence in 2003. What followed is more remarkable than the project itself: the cost of sequencing a human genome fell by orders of magnitude within roughly a decade, at a rate exceeding semiconductor cost decline over comparable periods.
The mechanism was a change in method. Chain-termination sequencing reads one fragment at a time, well and slowly. Massively parallel methods immobilise enormous numbers of fragments, read them simultaneously in short lengths, and reassemble the result computationally. Accuracy per read is lower and read length is shorter; both are compensated by reading each region many times over and by heavy computation.
The winning method is worse at the individual measurement and overwhelmingly better in aggregate, because the individual measurement was cheap enough to repeat massively. This trade — accept a poorer element in exchange for parallelism and then recover quality statistically — recurs in detector arrays, distributed computing, sensor fusion and manufacturing inspection. It is worth checking for whenever a process is limited by the cost of doing one thing well.
The consequence is that the bottleneck moved. Generating sequence data stopped being the constraint; interpreting it became one, and storing and moving it became another. Relieving a bottleneck relocates it, which this series has now observed in shipping, in networks and here.
05Takeaways for current practice
- Recognise when you are solving an inverse problem. If the instrument measures integrals or projections, sampling and conditioning matter more than detector quality.
- Make a measurable quantity encode the one you want. Gradients turn frequency into position; the same substitution trick is widely available.
- A worse element can win on parallelism. Cheap, repeatable, imperfect measurements beat expensive precise ones once statistics can recover the quality.
- Understand how an image was produced before trusting it. Artefacts are consequences of acquisition and inversion, not equipment faults.
- Expect the constraint to move. Cheap data generation made interpretation and storage the new limits.