Pushing the limits of spatial resolution in clinical PCD-CT using a dedicated high-resolution convolutional neural...

Area of Science:

• Medical Imaging Technology
• Artificial Intelligence in Radiology
• Photon-Counting Detector CT

Background:

• Photon-counting-detector (PCD) CT systems offer superior spatial resolution but clinical images are often limited by large pixel sizes and increased noise with sharp kernels.
• Current imaging protocols do not fully leverage the intrinsic high-resolution capabilities of PCD-CT systems.

Purpose of the Study:

• To evaluate the impact of pixel size and reconstruction kernels on visual spatial resolution in PCD-CT.
• To develop a high-resolution deep convolutional neural network (HR-CNN) to enhance spatial resolution in clinical PCD-CT imaging.

Main Methods:

• Investigated the relationship between spatial resolution, reconstruction kernel, and pixel size using phantoms and clinical data.
• Developed and trained a HR-CNN using ultra-high-resolution PCD-CT data (150-mm FOV, 0.15-mm pixel size, Qr89 kernel).
• Evaluated image quality, noise, and visual spatial resolution across four conditions, including HR-CNN, with two radiologists.

Main Results:

• Reducing field of view (FOV) to 150 mm with a sharp kernel (Qr89) improved resolution to 18-20 lp/cm.
• The HR-CNN further enhanced resolution to 20-22 lp/cm and was ranked superior by radiologists for image quality and spatial resolution.
• HR-CNN achieved significant noise reduction (93.0% vs. FBP, 44.9% vs. IR).

Conclusions:

• Routine PCD-CT practice often underutilizes its spatial resolution potential due to large FOV and noise.
• The HR-CNN denoising method, coupled with small pixel sizes, enables practical implementation of near-system-limit spatial resolution.
• This advancement is beneficial for diagnosing various conditions, including interstitial lung disease.