PKAID-Net: Prior Knowledge Aware Iterative Denoising Neural Network for Photon Counting Detector CT

Abstract

An important feature enabled by Photon-Counting Detector (PCD) CT is the simultaneous acquisition of multi-energy data, which can produce virtual monoenergetic images (VMIs) at a high spatial resolution. However, noise levels observed in the high-resolution (HR) VMIs are markedly increased. Recent work involving deep learning methods has shown great potential in CT image denoising. Many CNN applications involve training using spatially co-registered low- and high-dose CT images featuring high and low image noise, respectively. However, this is implausible in routine clinical practice. Further, typical denoising methods treat each VMI energy level independently, without consideration of the valuable information in the spectral domain. To overcome these obstacles, we propose a prior knowledge-aware iterative denoising neural network (PKAID-Net). PKAID-Net offers two major benefits: first, it utilizes spectral information by including a lower-noise VMI as a prior input; and second, it iteratively constructs refined datasets for neural network training to improve the denoising performance. This study includes 10 patient coronary CT angiography (CTA) exams acquired on a clinical HR PCD-CT (NAEOTOM Alpha, Siemens Healthineers). The HR VMIs were reconstructed at 50 and 70 keV, using a sharp kernel (Bv68) and thin (0.6 mm, 0.3 mm increment) slice thickness. Results showed that the PKAID-Net provided a noise reduction of 96% and 70% relative to FBP and iterative reconstruction, respectively while maintaining spatial and spectral fidelity and a natural noise texture. These results demonstrate the noise reduction capacity of PKAID-Net as applied to cutting-edge PCD-CT data to enable HR, multi-energy cardiac CT imaging.

1. INTRODUCTION

High-resolution (HR) PCD-CT can produce VMIs at multiple energy levels from the same scan data that have similar structural information but different spectral information and noise levels. This provides new opportunities for many clinical areas, such as cardiac CT, which can benefit from both high spatial and temporal resolution and multi-energy capabilities, especially in challenging scenarios involving the assessment of patients with dense calcifications and/or coronary stents. However, the increased image noise observed in HR VMIs limits the clinical adoption of this technique. Data redundancy in the spectral domain could be exploited as prior knowledge and enable improved VMI image quality via noise reduction techniques ¹. Recent deep learning-based methods have achieved great advancements in medical image denoising^2–6; however, the lack of large-scale patient datasets makes training deep learning models in clinical scenarios difficult. In addition, standard denoising methods treat each VMI energy level independently, without consideration of the valuable information in the spectral domain. To overcome these limitations, an image-based prior knowledge-aware iterative denoising neural network (PKAID-Net) framework was developed to iteratively create refined datasets for training a better denoising convolutional neural network (CNN). In this framework, we treat a lower-noise VMI as a prior input channel and update the training targets to improve the trained CNN denoising performance. The data preparation required for PKAID-Net module training involves four key processes: 1) spatial decoupling of the “noise-only” image maps, which mitigates overfitting and improves randomization, 2) slice averaging to produce a thicker (lower-noise) slice for the “signal-only” and “prior” images, 3) using thick reference images as training inputs, with reinsertion of spatially 1decoupled noise-only images and the refined prior images, and 4) using training targets that consist of the corresponding thick reference images without noise insertion. In this work, the denoised image from the previous training iteration was used as an updated signal-only target image, which was included in the dataset for the next training iteration. This process was implemented iteratively and was observed to gradually train high-performing denoising networks. Finally, the state-of-the-art denoising CNN was deployed from the PKAID-Net when a similar denoising performance compared to the trained CNN from last iteration is achieved (i.e., the standard deviation (STD) of the pixel values becomes stable).

2. METHODS

2.1. Overall Workflow of PKAID-Net

An overall illustration of our proposed PKAID-Net can be found in Figure 1(a). Given a series of PCD-CT VMI data, PKAID-Net first divides them into two inputs: the high-noise HR VMI and the prior VMI. Based on previous studies ^{7, 8}, 70 keV VMI was observed to have the lowest noise among all VMI energies, hence was appropriate for use as the prior VMI data input in this study. The characteristics of increased image contrast and decreased calcium blooming can be found in 50 keV and 100 keV images, while the noise is higher in both images. The 50 keV or 100 keV VMI reconstructed by the FBP method was chosen as the noisy VMI input. For the initialization of creating the dataset, the iterative reconstruction (IR) method is first applied to the noisy image to generate the initialized signal-only image. The noisy, signal-only, and prior images are then fed into the PKAID-Net Module to train a CNN. At each iteration m ( m = 1, …, M), where M represents the total number of iterations, the denoised image from the previous trained CNN-m-1 is used as a new signal-only target of the dataset for the next round of PKAID-Net Module-m training. A well-trained CNN will be obtained until a stable denoising performance is achieved during the iterations. For the inference of each trained CNN, the PCD-CT data with the noisy FBP and prior images are used as the inputs as shown in Figure 1(c), More details of the PKAID-Net Module are introduced in the remaining part of this section.

Figure 1.

Overview of the proposed Prior Knowledge Aware Iterative Denoising Neural Network (PKAID-Net). (a) The overall pipeline of PKAID-Net. (b) Overview of training a PKAID-Net Module. (c) The simplified inference process via the well-trained CNN-M from the PAKID-Net Module.

2.2. PKAID-Net Module

As shown in Figure 1(b), in the PKAID-Net Module for training a CNN, the noisy and signal-only images were first subtracted to generate images composed predominately of noise, called noise-only images. Spatial decoupling on the noise-only image was used to avoid overfitting and improve the randomization, which was defined as a random translation in the axial plane (1 to 16 pixels) and a random inversion (1 or −1 multiplier) of the noise-only image. A random noise image can be obtained with the above process. Meanwhile, five adjacent signal-only images and prior images were averaged to simulate thicker slice reference images as the refined signal-only and prior images with relatively low noise. Patch extractions were applied to these available images (Noise, refined signal-only, and prior images) to create the training dataset. The training target consisted of the refined signal-only patch without added noise. The inputs to training a CNN consisted of two channels (1) noise-only images multiplied by the weighting factor superimposed onto the refined signal-only patches and (2) prior patches. Denoising strength was controlled by the introduced weighting factor of the noise patch. To this end, a CNN can be trained with the created dataset as an output of the PKAID-Net Module.

2.3. CNN architecture in each module

A simplified U-Net architecture with nine modules was used in this study as shown in Figure 2. Each module involves convolution, batch normalization (BN), and exponential linear unit (eLU) activation operations sequentially. The max pooling layer and convolution transpose operator are applied in the network. The concatenation is added to the network to preserve the similarity between the input and output.

Figure 2.

The simplified architecture of the used U-Net model.

2.4. Dataset and training

Coronary CTA exams of 10 patients acquired on a commercial PCD-CT (NAEOTOM Alpha, Siemens Healthineers) were included in this study following approval by our Institutional Review Board. Scans were performed using the following parameters: prospective ECG-gated adaptive sequential mode, 120 kV, 144 × 0.4 mm collimation, and 0.25 second rotation time. Automatic exposure control (AEC) was turned on, with CAREkeV optimized for vascular exams and an Image Quality (IQ) level of 32. VMIs at 50 and 70 keV were reconstructed with both FBP and IR (strength level 4), with 1024 × 1024 matrix, Bv68 kernel, 0.6 mm slice thickness and 0.3 mm increment.

These patient images were divided into training (6 patients), validation (1 patient), and testing (3 patients) datasets. In this study, 50 keV images were used as the noisy VMI input and 70 keV images processed with a CNN-based denoising algorithm ⁹ were used as the prior. For the network training of each PKAID-Net Module, we randomly extracted 144,000 patches with the size of 128 × 128 pixels from the training data for each image type (noisy, signal-only, and prior) and 16,000 patches from the validation data by the ratio of 9:1. The initial learning rate was set as 0. 001 with a scheduled descent to 0.00001. And the Adam optimizer was chosen to minimize the mean-squared-error (MSE) loss function.

3. RESULTS

Figure 3 shows a representative slice from the 50 keV VMI as observed with different algorithms. Considering the PCD-CT images using a Bv68 sharp kernel and 0.6 mm slice thickness, our proposed PKAID-Net exhibits better visual quality in terms of detail preservation and noise removal, compared to the standard IR. Images from PKAID-Net (M=1) represent the output of 1^st iteration of the model, which is equivalent to a non-iterative version of the CNN denoising. Substantial noise reduction has been achieved in these images, more than that of IR. Further noise reduction and detail enhancement was observed at higher iterations, e.g., PKAID-Net (M=3). The red arrow in figure 3 indicates that the PKAID-Net (M=3) better preserves structural edges and maintains spatial resolution by comparison to the other variants. The proposed method was evaluated quantitatively by measuring the mean and standard deviation (STD) values within a region of interest (ROI) placed on the aorta (red circle in Figure 3). Values are summarized in Table 1. It is indicated that our PKAID-Net (M=3) can remove 96% (41 vs 1292 HU) and 70% (41 vs 154 HU) noise relative to the FBP and IR method, respectively. The performance of different iterations for the PKAID-Net was compared and the results demonstrated that optimal performance can be achieved with a small number of iterations (M=3), after which the difference became negligible. Additionally, the mean attenuation value was similar to that of the FBP and IR methods, which verifies the robustness of the PKAID-Net at maintaining the CT numbers and spectral information (as CT numbers of VMIs represent the spectral property of the images).

Figure 3.

Denoising performance comparison for a representative slice of the 50keV VMI data f rom one patient. I mages f rom columns left to right are FBP, IR, PKAID-Net (M=1), and PKAID-Net (M=3), where M represents the total number of iterations. The regions of interest marked by the yellow rectangle are zoomed below, respectively. Image display window (WW/WL): 400/1500 HU.

Table 1. DENOISING PERFORMANCE.

	FBP	IR	PKAID-Net (M=1)	PKAID-Net (M=2)	PKAID-Net (M=3)
Mean	1217	1225	1215	1204	1216
Std	1192	154	52	42	41

comparison of between FBP, IR, and PKAID-Net, Unit: HU.

4. CONCULUSION

A prior knowledge-aware iterative denoising neural network (PKAID-Net) scheme was developed to perform noise reduction in clinical PCD-CT applications. This denoising method is able to fully utilize PCD-CT image data in the spectral domain, allowing iterative refinement of training datasets. The image domain denoising enables easy deployment upon clinical scans. The experimental results testing on patients’ cardiac CTA data demonstrated that the proposed PKAID-Net enables substantial noise reduction while maintaining spatial and spectral fidelity of the HR PCD-CT in clinical exams.