Cherenkov Luminescence Microscopy: A network trained on Monte Carlo simulations cleans images of tumour spheroids

Driven on 15,360 pairs of simulated images, a CARE network divides the reconstruction error against Noise2Void by nearly ten, then improves the contrast on 27 real spheroids marked with 18F-FDG.

The luminescence microscopy Cherenkov shows where a radiotracer attaches to 3D tumour spheroids, models of tumours grown in the laboratory. But the images suffer from a small number of photons and blurry due to the journey of the positons before their annihilation. An Italian team proposes to restore them with a network of neurons trained not on real images, impossible to get without fault, but on physical simulations.

With TOPAS/Geant4, researchers modelled 80 heterogeneous spheroid geometries, simulating positon transport, Cherenkov emission and optical propagation, to produce 15,360 pairs of degraded and ideal images. The supervised CARE network thus led reduces the mean quadratic error from 0.0492 to 0.0051 on validation data, much better than the self-supervised Noise2Void method. Applied to experimental images of 27 U87 spheroids having captured 18F-FDG, it gives more localized distributions, with better contrast and better signal to noise ratio. The authors see this as a tool consistent with the 3R rule to reduce animal testing.