Ultra-High-Resolution Photon-Counting-Detector CT with a Dedicated Denoising Convolutional Neural Network for...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Ultra-high-resolution (UHR) photon-counting-detector (PCD) CT offers improved spatial resolution but suffers from increased image noise.
• Smoother reconstruction kernels are typically used to mitigate noise, but this compromises the achievable resolution.
• This limitation hinders optimal visualization of fine temporal bone anatomy.

Purpose of the Study:

• To develop and evaluate a denoising convolutional neural network (CNN) for UHR PCD-CT.
• To reduce noise in images reconstructed with the sharpest kernel, preserving high resolution.
• To enhance the visualization of temporal bone structures for improved diagnostic accuracy.

Main Methods:

• A CNN was trained on 1885 images from 6 patient cases and tested on 20 independent cases using a dual-source PCD-CT (NAEOTOM Alpha).
• Images were reconstructed using quantum iterative reconstruction (QIR3) with both a routine (Hr84) and the sharpest (Hr96) kernel.
• The CNN was applied to Hr96-reconstructed images (QIR1), and results were compared to conventional reconstructions (Hr84-QIR3, Hr96-QIR3) by neuroradiologists.

Main Results:

• The CNN reduced noise by 80% compared to Hr96-QIR3 and 50% compared to Hr84-QIR3, while maintaining high resolution.
• Noise levels decreased from 204.63 to 47.35 HU with Hr96-CNN compared to Hr96-QIR3.
• Image quality and visualization of the modiolus, stapes footplate, and incudomallear joint were significantly preferred with the CNN-processed images.

Conclusions:

• The developed CNN effectively reduces noise in UHR PCD-CT images.
• This enables the use of the sharpest reconstruction kernels without compromising image quality.
• The combination of the CNN and sharpest kernel significantly enhances diagnostic image quality and anatomic visualization of the temporal bone.