Exploring charge sharing compensation using inter-pixel coincidence counters for photon counting detectors by...

Area of Science:

• Medical Imaging
• Detector Physics
• Machine Learning

Background:

• Photon counting detectors (PCDs) offer advantages in computed tomography (CT) but face challenges like charge sharing, limiting their diagnostic potential.
• Multi-energy inter-pixel coincidence counters (MEICC) provide spatial information to address charge sharing, potentially lowering the Cramér-Rao Lower Bound (CRLB).

Purpose of the Study:

• To explore charge sharing compensation in MEICC detectors using a deep-learning method that leverages local spatial coincidence counter information.
• To evaluate the effectiveness of the deep-learning approach in improving virtual monochromatic attenuation integral (VMAI) estimation compared to conventional PCDs.

Main Methods:

• A deep-learning network was designed to focus on individual pixels, using MEICC data patches as input for charge sharing compensation.
• A fast, online data generation method and a novel loss function for high-noise data were developed.
• Validation was performed using Monte Carlo (MC) simulation data, comparing MEICC detectors with conventional PCDs.

Main Results:

• The deep-learning method achieved minimal bias (0.6-1.3%) and reduced standard deviation/NRMSE in VMAI estimation, outperforming polynomial fitting (>3% bias).
• MEICC detectors showed superior performance across all metrics, with approximately 10% lower noise compared to conventional PCDs.
• An ablation study confirmed the benefit of an additional loss function for high-noise data training.

Conclusions:

• Network-based methods can effectively utilize local information from PCDs for charge sharing compensation via patch-based learning.
• MEICC detectors offer valuable local spatial information, enabling more accurate VMAI estimation than conventional PCDs.
• The proposed deep-learning approach enhances diagnostic CT imaging by mitigating charge sharing artifacts.