PKAID-Net: Prior Knowledge Aware Iterative Denoising Neural Network for Photon Counting Detector CT

Area of Science:

• Medical Imaging
• Radiology
• Deep Learning

Background:

• Photon-Counting Detector (PCD) CT allows simultaneous multi-energy data acquisition, producing high-resolution virtual monoenergetic images (VMIs).
• High-resolution VMIs often exhibit increased noise levels, limiting their clinical utility.
• Existing deep learning denoising methods face challenges with clinical data acquisition and independent VMI energy level processing.

Purpose of the Study:

• To develop and evaluate a novel deep learning network, the prior knowledge-aware iterative denoising neural network (PKAID-Net), for denoising high-resolution VMIs from PCD-CT.
• To leverage spectral information and iterative dataset refinement for improved denoising performance.
• To address the limitations of current denoising techniques in clinical settings.

Main Methods:

• The proposed PKAID-Net utilizes a lower-noise VMI as a prior input to incorporate spectral information.
• The network iteratively refines training datasets to enhance denoising capabilities.
• The method was tested on 10 patient coronary CT angiography (CTA) exams from a clinical HR PCD-CT, reconstructing VMIs at 50 and 70 keV with a sharp kernel and thin slices.

Main Results:

• PKAID-Net achieved significant noise reduction: 96% compared to Filtered Back Projection (FBP) and 70% compared to iterative reconstruction.
• The denoising process maintained spatial and spectral fidelity of the images.
• A natural noise texture was preserved in the denoised images.

Conclusions:

• PKAID-Net demonstrates substantial noise reduction capacity for high-resolution PCD-CT data.
• The method effectively enables high-resolution, multi-energy cardiac CT imaging by overcoming noise limitations.
• PKAID-Net shows promise for advancing clinical applications of cutting-edge PCD-CT technology.