Enabling Low-Dose In Vivo Benchtop X-ray Fluorescence Computed Tomography through Deep-Learning-Based Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• X-ray Fluorescence Computed Tomography (XFCT) offers high-resolution, non-invasive molecular imaging.
• Current XFCT techniques require high radiation doses for adequate sensitivity, posing safety concerns.
• Artificial Intelligence (AI), particularly deep learning (DL), shows potential for noise reduction in medical imaging.

Purpose of the Study:

• To develop and evaluate an optimized Swin-Conv-UNet (SCUNet) deep learning model for background noise reduction in X-ray Fluorescence (XRF) images.
• To enable high-quality XFCT imaging from low-dose X-ray fluorescence data.
• To mitigate the trade-off between imaging sensitivity and radiation exposure in XFCT.

Main Methods:

• An optimized Swin-Conv-UNet (SCUNet) deep learning model was developed for background noise reduction.
• The model was trained and assessed using augmented XRF data, focusing on low tracer concentrations.
• Image quality was quantitatively evaluated using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) against high-dose images.

Main Results:

• The SCUNet model successfully generated high-quality XRF images from low-dose inputs, achieving a maximum PSNR of 39.05 and SSIM of 0.86.
• The proposed DL algorithm demonstrated superior performance compared to traditional denoising methods like BM3D, BM4D, NLM, and DnCNN, especially in high-noise conditions.
• Visual inspection and quantitative metrics confirmed the effectiveness of the SCUNet model in noise reduction.

Conclusions:

• The optimized SCUNet model effectively reduces background noise in XFCT imaging at low tracer concentrations.
• AI-driven denoising using SCUNet allows for high-quality XFCT imaging with significantly reduced radiation exposure.
• This approach holds substantial potential for advancing XFCT applications in molecular imaging.