A CT scan is many X-ray pictures turned into one 3D image. Take X-ray pictures of something from every side and a computer can work out what is inside it, slice by slice. On this machine the scan is also the map for treatment, because the beam can only hit what the scan shows, so the two must agree exactly on where everything is.
We added a reconstruction engine into the machine’s control software that runs on an NVIDIA GPU in the control computer and works with more than one make of flat-panel detector, so the same machine could ship with different hardware. With GPU reconstruction, the machine’s whole imaging step was reduced to about 1 minute.
Why the subject spins and the scanner stays still
Most CT scanners spin the X-ray source and detector around the patient, but here they stay still while the stage turns the subject a full circle between them. The same turning stage gives the treatment beams extra angles, and the beam’s center sits on the stage’s axis of rotation, so the scan and the treatment share one frame of reference.
Why NVIDIA GPUs
Reconstruction repeats the same small calculation for millions of points in the volume, which is the kind of work a graphics card is built for. On a GPU it drops from many minutes to seconds (one 2010 study: 25 minutes on a processor, 3.2 seconds on a GPU), and the licensed GPU reconstruction brought the machine’s whole imaging step down to about a minute.
The shipped system offered licensed reconstruction software alongside ours. Ours runs on a dedicated NVIDIA Tesla C2075 compute card in the control computer, using CUDA tools from Johns Hopkins’ I-STAR lab.