Denoising pediatric cardiac photon-counting CT data with sparse coding and data-adaptive, self-supervised deep...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pediatric Cardiology

Background:

• Pediatric cardiac CT imaging requires careful radiation dose management due to repeated scans and increased lifetime cancer risk.
• Image quality in pediatric cardiac CT is often variable due to protocol limitations, metallic implants, and denoising algorithm performance disparities.
• Photon-counting CT (PCCT) and deep learning (DL) offer advancements for improved pediatric CT scan quality at reduced radiation doses.

Purpose of the Study:

• To enhance self-supervised deep learning (DL) denoising techniques for pediatric cardiac CT data with variable image quality.

Main Methods:

• A modified 3D Vision Transformer (mViT) was developed, incorporating architectural changes for cross-token recombination and sparse coding.
• The mViT was trained dynamically, balancing data fidelity and representation sparsity based on local image noise estimates.
• Training utilized pediatric cardiac photon-counting CT data from 20 patients (ages 1-18) with varying noise levels.

Main Results:

• The mViT with sparse coding preserved diagnostic anatomical structures in denoised images, outperforming other methods in intensity variance.
• The trained network showed robust generalization to preclinical PCCT data with high noise levels and different contrast.
• Application to clinical PCCT data in infants (<1 year) revealed minor smoothing of details in already denoised images.

Conclusions:

• This work presents a robust, self-supervised denoising method for pediatric cardiac PCCT, adapting network training to local noise estimates.
• The trained network demonstrates generalization to diverse noise levels and contrast variations.
• Self-supervised fine-tuning suggests potential for addressing related CT denoising challenges.