Computing Cases Header, Picture of a Keyboard with the text "" printed over it

The Atomic Energy Canada, Limited and Therac-25

Early Therac Machines

The story of Therac-25 begins in the early 1970's when Atomic Energy Canada, Limited (AECL) joined forces with a French company, CGR, to design and build a medical linear accelerator based on earlier CGR machines. The companies cooperated on the design and manufacture of two successful medical linear accelerators, the Therac-6 and its successor, the Therac-20. Both these machines were based on CGR designs that did not use computer control. The new machines added computer control, in addition to other innovations. The Therac-6 was the initial product of their collaboration and was designed to produce X-rays for radiation therapy. The Therac-20 was a much more powerful and versatile machine. It could produce two different kinds of radiation beams for treatment of deep and shallow tissue. AECL also produce other medical linear accelerators, including the Therac-4, a single mode electron beam machine.

Development of Therac-25

In the early 1980's, AECL developed a much more space-efficient medical linear accelerator that was just as powerful and versatile as the Therac-20. Linear accelerators are more powerful the longer they are, and AECL found a way to fold the long beam-producing mechanism for a 25 MeV machine into a smaller space. In addition, this new version was somewhat less expensive to produce, since it used a less expensive beam production device (a magnetron instead of a klystron).

Finally, AECL intended to take advantage of increasing capability of computer software to make the machine easier to operate. The new Therac-25 was the result of a convergence of the new beam-folding technology with the ease of computer control, bringing with it the bonus of lower production costs. In addition to lower production costs, the computer control allowed faster setup of the machine for each patient. This meant that more patients could be treated in one day than with non-computerized linear accelerators.

The Therac-25's ancestors, Therac-20 and Therac-6, had used a minicomputer (a DEC PDP-11) to add some convenience to the standard hardware of a medical linear accelerator. They both could work without computer control. AECL determined to make its new model, Therac-25, a tightly-coupled combination of software and hardware. By this time, its collaboration with CGR had grown stale and AECL was bringing in its new beam folding technology (and the new Therac-25) on its own.

In tightly coupling the software and the hardware, AECL could use the software to monitor the state of the machine for proper operation and for safety. Previous versions, with designs based in models that predated computer control, had included independent circuits to monitor beam scanning and had mechanical interlocks to ensure the machine could not enter a state in which it could harm a patient. But with increased computer control, AECL decided not to duplicate this equipment in the Therac-25 (with some cost savings), and to rely on software for policing these safety issues.

Therac-25 goes to Market

In late 1982, Therac-25 was first offered to hospitals in a commercial version. It was eventually adopted by eleven institutions, six in Canada and five in the US. These included sites in Georgia, Texas, Washington State, and Hamilton, Ontario.

Safety Analysis of Therac-25

In 1983, just after AECL made the Therac-25 commercially available, AECL performed a safety analysis of the machine using Fault Tree Analysis. This involved calculating the probabilities of the occurrence of varying hazards (e.g. an overdose) by specifying which causes of the hazard must jointly occur in order to produce the hazard.

Since much of the software had been taken from the Therac-6 and Therac-20 systems, and since these software systems had been running many years without detectable errors, the analysts assumed there were no design problems in the software. The analysts did consider software failures like "computer selects wrong mode" but assigned them probabilities like 4 x 10**-9. These sorts of probabilities are likely assigned based on the remote possibility of random errors produced by things like electromagnetic noise. They do not take into account the possibility of design flaws in the software.