XCT2PCCT: Accurate Simulation of in-vivo Photon-Counting Computed Tomography of the Human Spongiosa from High Resolution Peripheral Computed Tomography

FELIX SEBASTIAN LEO THOMSEN; JAN BORGGREFE

Minden, Alemania

Conferencia; Wissenschaftstag Mühlenkreiskliniken 2022; 2022

Ruhr-Universität Bochum

Development of new methods for the analysis of bone microstructure improves the diagnosis of osteoporosis and other structural bone pathologies. Machine learning (ML) methods such as convolutional neural networks, adversarial generative networks etc., are most promising here. Employment of ML however requires a large size of training data to obtain robust performance for generalizing the learned knowledge for real images. In particular for bone microstructure the obtainment of a sufficient large training set is very expensive and impractical, which typically requires pairs of real ground-truth high-resolution peripheral (XCT) and clinical CT data of the same specimen with a sufficient variability of ray doses, several repetitions and possibly different reconstruction kernels and resolutions.This study aims for a deterministic method to simulate a wide range of settings of microstructural in-vivo Photon-Counting CT (PCCT). We employ kernels Br64 - Br76, all iterative reconstruction levels QIR 0 to 4, exposures between 60 and 375 mAs, slice increments between 0.1 and 1.0 mm, the Ultra-Quantum-Plus (ultra-high-resolution UHR) and Quantum-Plus mode (high-resolution HR and standard-resolution SR). Simulation of the PCCT signal has been performed with the noise power spectrum (NPS), the optical transfer function (OTF) and the radial anisotropy of the bone signal and noise. In combination with a generative adversarial network (GAN) for artificial XCT scans, trained for the same resolution and patch size, this method allows to generate an infinite set of bone training data for the development of new osteoporosis markers and thereby to circumvent the conduction of a large number of costly physical scans.