Polychromatic neural CBCT reconstruction through density-attenuation modeling

Monochromatic cone beam computed tomography reconstruction algorithms are still most prominent in practice. Since the x-ray detectors of today's machines are mostly energy integrating detectors and thus not able to resolve photon energy levels, reconstruction artifacts, often termed as beam-hardening artifacts, are a frequent observation. We propose a polychromatic 3D reconstruction technique, using a coordinate based neural representation, that requires zero additional prior information, in which we model attenuation related to a fixed energy levelE₀together with its derivative. We concurrently optimize an intermediate density value for each spatial position together with a composite attenuation function. It assigns each intermediate density value its attenuation coefficient and computes the derivative of the attenuation coefficient w.r.t. the energy. We implement this function as a neural network that models a monotonic relationship between intermediate density and attenuation. In our results we can show that the proposed method is able to considerably improve reconstruction quality in various scenarios, both quantitatively using a synthetic numerical phantom as well as visual quality in real-world clinical examples.