You taste a fruit smoothie through two different straws. One lets through only the big chunks, the other only the liquid. Individually, the sensations are misleading, but together they give you the key: here, it seems, is banana, and here, strawberry. That's roughly how dual-energy computed tomography works: it shines X-rays through the body at two different energies, then separates the signal into specific materials — for example, telling bone from muscle or spotting a drop of contrast agent in a vessel.
This separation is an inverse problem: you need to recover the addends from the sum, and the slightest noise in the data turns into ripples on the final images. Conventional methods either compute slowly or require perfect reference images, which are almost never available in a real clinic. The authors of the new study proposed training a deep learning model without a single ground truth. Instead of showing it pairs of "raw signal — clean image," they embedded the physical model of the scanner itself into the training. The network makes a guess about tissue composition, passes it through a virtual X-ray machine, and compares the result with actual measurements. If the discrepancy is large, it corrects its guess, like an artist who constantly checks against the subject rather than copying someone else's drawing.
The method was tested on open data from the AAPM DL-Spectral CT challenge. In terms of accuracy in recovering the original projections, it outperformed three modern algorithms. But the most interesting part is the virtual monoenergetic images. It's as if from the same smoothie you could get the taste of each fruit separately, and at different "temperatures": at 30, 50, and 70 keV. At all three levels, the new approach produced fewer errors and sharper image structure than the competitors. That means a doctor could more confidently spot a tumor boundary or a tiny clot — all without an extra radiation dose.
Dual-energy CT can compute the concentration of iodine in tissues — which is almost the same as measuring blood flow in a tumor without injecting an additional dye.
In the series 'The Expanse', medical scanners instantly determine the composition of any object by its spectral signature — our method takes a step toward such diagnostics without reference samples.