Simulating and correcting the pileup effect in deep-silicon photon-counting CT

Photon-counting computed tomography (CT) bears promise to substantially improve spectral and spatial resolution. One reason for the relatively slow evolution of photon-counting detectors in CT-the technology has been used in nuclear medicine and planar radiology for decades-is pulse pileup, that is, the random staggering of pulses, resulting in count loss and spectral distortion, which in turn cause image bias and reduced contrast-to-noise ratio (CNR). The deterministic effects of pileup can be mitigated with a pileup-correction algorithm, but the loss of CNR cannot be recovered, and must be minimized by hardware design. In the deep-silicon photon-counting detector, each pixel is split into depth segments, which enables optimization of the count rate per detector channel to reduce pileup. Virtual clinical trials are attracting growing interest for efficient evaluation of cutting-edge technology like the deep-silicon design, but a virtual trial requires an accurate simulation model of the imaging system, a digital twin, which captures all relevant aspects of the system over the full spectrum of clinical applications.

We are developing a framework for digital twins of deep-silicon photon-counting CT to enable in-silico system evaluation and virtual clinical trials of the technology. The primary purpose of this study is to validate the framework with respect to pileup, that is, it is not a validation of the detector performance, but a validation of the correspondence between simulation and measurements from a prototype device. A secondary purpose is to employ the framework for investigating the impact of pileup on image quality and the effectiveness of a data-driven pileup correction algorithm.

A pileup model that simulates individual photon events in accordance with the semi-nonparalyzable detector behavior was integrated into the CatSim environment. Measured count data from a prototype deep-silicon system were used to validate the simulation framework with respect to pileup. A typical image chain was integrated into the framework, including material decomposition (MD) and data-driven pileup correction. Images of a software phantom were generated to illustrate the effect of pileup on images and to assess the effectiveness of the pileup correction algorithm.

Simulated data were described well by the semi-nonparalyzable detector model and exhibited deviations to the measured count rate and variance of less than 5% across energy bins and depth segments, and a wide range of tube currents. The investigated pileup correction algorithm suppressed artifacts to below the noise level in monochromatic images and material images, and reduced iodine bias from 26% to 2% in the range from a factor of 3 lower to a factor of 1.7 higher than the calibrated count rate without impacting CNR.

The observed discrepancies are reasonable given known uncertainties, and the model provides a reliable representation of the pileup effect. The framework for digital twins helped confirm adequate performance of the pileup correction algorithm, which can reduce the need for repeated MD calibrations in mA-modulated scans. Next steps include simulation speed up and expansion of the framework to other detector effects.