Spectral deep learning-based patient and bowtie scatter correction for clinical photon-counting CT

Area of Science:

• Medical Imaging
• Computed Tomography (CT)
• Photon-Counting Detector Technology

Background:

• Scatter in CT degrades image quality, originating from patient tissues and beam-hardening filters (e.g., bowtie filters).
• Conventional energy-integrating detectors lack spectral distinction, unlike photon-counting (PC) detectors which offer energy-selective measurements.
• PCCT spectral data inherently contains information about scatter, as different energy thresholds are affected uniquely by scatter.

Purpose of the Study:

• To investigate the exploitation of spectral information from PCCT for enhancing deep learning (DL)-based scatter correction.
• To evaluate the performance of joint versus separate correction of patient and bowtie scatter using DL approaches.
• To address the under-consideration of bowtie scatter in existing DL-based scatter correction methods.

Main Methods:

• Development of a DL-based approach for joint estimation of patient and bowtie scatter, compared against separate correction methods.
• Introduction of neural network architectures incorporating PCCT spectral information for scatter correction, capable of estimating scatter across up to four energy thresholds.
• Validation using both Monte Carlo simulated data and real-world data acquired from a clinical PCCT system.

Main Results:

• Both joint and separate scatter estimation methods reduced Mean Absolute Error (MAE) from 8 HU to 1 HU.
• Spectral Deep Scatter Estimation (DSE) methods significantly outperformed non-spectral and convolution-based approaches, reducing scatter errors from up to 8 HU to below 1 HU across all energy thresholds.
• Spectral DSE with four energy thresholds achieved the best performance, reducing critical MAE (MAE10) from 23.8 HU to 1.6 HU, significantly improving artifact intensity in the most affected image regions (25% of volume).

Conclusions:

• A unified DL method effectively corrects both patient and bowtie scatter, eliminating the need for multiple specialized networks and reducing computational complexity.
• Leveraging spectral information in DL-based scatter estimation substantially improves correction accuracy, particularly beneficial for spectral applications like virtual monoenergetic images (VMIs).
• The proposed spectral DSE approach, especially with multiple energy thresholds, enables more accurate scatter estimation and correction in PCCT.