X-ray fluorescence CT (XFCT) is a novel imaging modality that combines X-ray CT with X-ray fluorescence analysis (XRFA) and can be employed to probe the distribution and concentration of functionalized gold nanoparticles inside the tumor. It has good potential in the diagnosis and treatment of early-stage cancers. How to suppress Compton scattering noise for XFCT imaging is a current hotspot. Traditional denoising methods include the background fitting method, scanning phase subtraction, and iterative denoising method. Deep learning-based denoising and reconstruction methods can utilize the powerful feature learning ability of deep learning without priori information such as parameters of imaging systems, which can effectively reduce the background noise and obtain sound imaging quality.

We propose an XFCT denoising algorithm based on noise level estimation and convolutional neural networks (NeCNN), which consists of noise estimation subnetworks and main denoising networks (Fig. 2). The estimated subnetwork estimates the noise level and reduces the preliminary noise through the denoising convolutional neural network (DnCNN). The estimated results are input into the fully convolutional neural network (FCN) and the output is adopted to learn the Compton scattering distribution. Meanwhile, as the FCN integrates a deconvolution module, the denoising and reconstruction of end-to-end fluorescence CT images can be directly achieved. We utilize the air-loaded phantoms for pre-training, while the related parameters are transferred into the PMMA phantoms to simulate the human tissue and achieve faster convergence. This two-level network structure is not a simple cascade, and the input-output and hyper parameter settings between two-level networks are linked to each other. With preliminary noise level estimation and input into the secondary network as priori information, there is a superior denoising effect compared with a single denoising network. Additionally, the mean square error (MSE) and structure similarity (SSIM) are employed as the loss function to get the local and global optimal solutions.

The imaging system contains an X-ray source, a phantom to be measured, two sets of pinhole collimators, and two sets of fluorescence detectors (Fig. 1). The distances between the fan beam X-ray source and the phantom center, between the pinhole collimator and the phantom, and between the detector and pinhole collimator are 15, 5, and 5 cm respectively. The detector consists of 55 × 185 cadmium telluride (CdTe) detector units with an energy resolution of 0.5 keV, and the crystal size is designed to be 0.3 mm×0.3 mm. The datasets are obtained with Geant4 software by scanning air phantom and PMMA phantom in which various metal nanoparticles (Au, Bi, Ru, Gd) are filled, and different incident X-ray intensities are set to simulate different noise levels and enhance the model's generalization ability. The imaging phantom is set as a cylinder with a diameter of 3 mm and a height of 5 cm, and the settings of element concentration are divided into two types, including high mass fraction versus low mass fraction, where high mass fraction includes 0.2%, 0.4%, 0.6%, 0.8%, 1.0%, and 1.2%, and low mass fraction includes 0.1%, 0.12%, 0.14%, 0.16%, 0.18%, and 0.2%. The programming language is Python 3.6 and the NeCNN is implemented based on Pytorch 1.7.0. Meanwhile, the hardware platform is configured as Intel i5-9600kf CPU, NVIDIA Titan V (12 GB/NVIDIA) GPU, and 16 G DDR4 RAM. The hyper parameters are shown in Table 1. Figure 6 shows the denoised images with NeCNN, BM3D, and DnCNN algorithms. We can easily find that both NeCNN and DnCNN can effectively reduce noise in the background region, which is difficult to handle for the BM3D algorithm. Additionally, NeCNN is more effective than DnCNN in removing abnormal pixel spots caused by self-absorption in the center region of interest (ROI). Generally, the proposed NeCNN is quantitively and qualitatively superior to the traditional BM3D and DnCNN algorithms. The NeCNN algorithm has the largest PSNR (29.01558) and SSIM (0.95066) values. Compared with DnCNN, NeCNN shows an improvement of 0.23993 and 0.02734 in terms of PSNR and SSIM respectively.

This sduty proposes a novel denoising algorithm for XFCT images based on deep learning to estimate the Compton scattering noise level by noise estimation subnetworks and noise reduction by the denoising main network. The experimental results show that for both air and PMMA phantoms, the PSNR and SSIM of images with NeCNN are both higher than DnCNN and BM3D. This illustrates the effectiveness of the proposed algorithm and shows its potential to be applied in practical imaging systems in the future.