An X-ray tomograph is like a carpenter who examines a log from all sides to see where the void is inside. It takes dozens of flat shadow images, rotating the source and detector around the object. Then a 'reconstruction' program glues these shadows into a 3D model. The quickest reconstructor is the FBP algorithm (filtered back projection). It works almost instantly, but there's a catch: when there are few scans or the radiation dose is reduced, the shadows become blurry. And then FBP starts smearing noise all over the image. Instead of a void in the log, you see ripples.
The authors proposed a different 'reconstructor'—a probabilistic one. Instead of just adding up shadows, it estimates: what internal structure of the object would most likely produce exactly these blurred projections? It is based on a Markov random field model — a way to describe how neighboring points in the volume influence each other. If neighboring voxels (volumetric pixels) are almost always similar in density, the algorithm gets a hint: a single bright point among dark ones is more likely noise than a real feature. The program tunes its own sensitivity by estimating the noise level directly from the input data — through minimization of Bayesian free energy. No manual adjustments.
Tests on simulations showed that under low-dose and few-angle conditions, the new method produces a much cleaner image than FBP. This opens the door to tomography where one previously had to choose between quality and safety: quality control of microelectronics without prolonged exposure, baggage screening with just a couple of angles, and medical diagnostics with reduced radiation load. Fast data acquisition and quiet, thoughtful reconstruction — instead of haste that ruins the result.
The word 'tomography' comes from the Greek 'tomos'—a slice, a layer. The first commercial CT scanner for the head appeared in 1971 and took one image in several minutes.