Room: AAPM ePoster Library
Purpose: The aim of this study is to develop a method to estimate an x-ray spectrum in cone-beam computed tomography (CBCT) systems from the percentage depth dose (PDD) by use of machine learning (ML) approaches.
Methods: We assume that the observed PDD is expressed by a weighted mean of monochromatic PDDs (mPDDs), which are formed by monochromatic x-ray energies. The mPDDs of monochromatic x-ray energies from 10keV to 140keV with a 5-keV interval were prepared by Monte Carlo (MC) simulation without details of head structure. With mPDDs, the x-ray spectrum prediction model was constructed with two different ML approaches, maximum a posterior model including a prior information of the continuous relationship between the adjacent energy bins and a generative model based on an artificial neural network. These two models were evaluated by the comparison between the estimated weights and the those calculated by full MC simulations in X-ray Volumetric Imager (XVI) system (ELEKTA) and On-Board Imager (OBI) system (Varian). For applications, we estimated the x-ray spectrum from the actual experimental PDD with nominal x-ray energies of 100kV and 120kV measured in XVI system.
Results: The x-ray spectrum estimated from PDD obtained with MC simulation was accurately reproduced both in XVI and OBI system. The average root-mean-squared error for the x-ray spectrum compared with MC simulation in the practical CBCT scanners was approximately 5×10?³ both in ML approaches. For the use of actual measured PDD in XVI system, the reproduction of the PDD was successfully performed in 100kV. However, an anomaly was found in 120kV and the reproduction become reasonable when the monochromatic x-rays with 130keV and 140keV were included.
Conclusion: This research showed that an x-ray spectrum estimation using ML approaches was feasible from a PDD measurement which does not require a complicated device and measurement.