Generative Adversarial Networks for Noise Reduction in Low-Dose CT.

Computer algorithms that simulate lower-doses computed tomography (CT) images from clinical-dose images are widely available. However, most operate in the projection domain and assume access to the reconstruction method. Access to commercial reconstruction methods may often not be available in medical research, making image-domain noise simulation methods useful. However, the introduction of non-linear reconstruction methods, such as iterative and deep learning-based reconstruction, makes noise insertion in the image domain intractable, as it is not possible to determine the noise textures analytically. To develop a deep learning-based image-domain method to generate low-dose CT images from clinical-dose CT (CDCT) images for non-linear reconstruction methods. We propose a fully image domain-based method, utilizing a series of three convolutional neural networks (CNNs), which, respectively, denoise CDCT images, predict the standard deviation map of the low-dose image, and generate the noise power spectra (NPS) of local patches throughout the low-dose image. All three models have U-net-based architectures and are partly or fully three-dimensional. As a use case for this study and with no loss of generality, we use paired low-dose and clinical-dose brain CT scans. A dataset of paired scans was retrospectively obtained. All images were acquired with a wide-area detector clinical system and reconstructed using its standard clinical iterative algorithm. Each pair was registered using rigid registration to correct for motion between acquisitions. The data was randomly partitioned into training ( samples), validation ( samples), and test ( samples) sets. The performance of each of these three CNNs was validated separately. For the denoising CNN, the local standard deviation decrease, and bias were determined. For the standard deviation map CNN, the real and estimated standard deviations were compared locally. Finally, for the NPS CNN, the NPS of the synthetic and real low-dose noise were compared inside and outside the skull. Two proof-of-concept denoising studies were performed to determine if the performance of a CNN- or a gradient-based denoising filter on the synthetic low-dose data versus real data differed. The denoising network had a median decrease in noise in the cerebrospinal fluid by a factor of and introduced a median bias of HU. The network for standard deviation map estimation had a median error of HU. The noise power spectrum estimation network was able to capture the anisotropic and shift-variant nature of the noise structure by showing good agreement between the synthetic and real low-dose noise and their corresponding power spectra. The two proof of concept denoising studies showed only minimal difference in standard deviation improvement ratio between the synthetic and real low-dose CT images with the median difference between the two being 0.0 and +0.05 for the CNN- and gradient-based filter, respectively. The proposed method demonstrated good performance in generating synthetic low-dose brain CT scans without access to the projection data or to the reconstruction method. This method can generate multiple low-dose image realizations from one clinical-dose image, so it is useful for validation, optimization, and repeatability studies of image-processing algorithms.

Deep learning has recently attracted widespread interest as a means of reducing noise in low-dose CT (LDCT) images. Deep convolutional neural networks (CNNs) are typically trained to transfer high-quality image features of normal-dose CT (NDCT) images to LDCT images. However, existing deep learning approaches for denoising LDCT images often overlook the statistical property of CT images. In this paper, we propose an approach to statistical image restoration for LDCT using deep learning. We introduce a loss function to incorporate the noise property in image domain derived from the noise statistics in sinogram domain. In order to capture the spatially-varying statistics of CT images, we increase the receptive fields of the neural network to cover full-size CT slices. In addition, the proposed network utilizes z-directional correlation by taking multiple consecutive CT slices as input. For performance evaluation, the proposed networks are trained and validated with a public dataset consisting of LDCT-NDCT image pairs. We also perform a retrospective study by testing the networks with clinical LDCT images. The experimental results show that the denoising networks successfully reduce the noise level and restore the image details without adding artifacts. This study demonstrates that the statistical deep learning approach can restore the image quality of LDCT without loss of anatomical information.

Coronary artery calcification (CAC) measurement from low-dose CT images can be used to assess the risk of coronary artery disease. A fully automatic algorithm to detect and measure CAC from low-dose non-contrast, non-ECG-gated chest CT scans is presented. Based on the automatically detected CAC, the Agatston score (AS), mass score and volume score were computed. These were compared with scores obtained manually from standard-dose ECG-gated scans and low-dose un-gated scans of the same patient. The automatic algorithm segments the heart region based on other pre-segmented organs to provide a coronary region mask. The mitral valve and aortic valve calcification is identified and excluded. All remaining voxels greater than 180HU within the mask region are considered as CAC candidates. The heart segmentation algorithm was evaluated on 400 non-contrast cases with both low-dose and regular dose CT scans. By visual inspection, 371 (92.8%) of the segmentations were acceptable. The automated CAC detection algorithm was evaluated on 41 low-dose non-contrast CT scans. Manual markings were performed on both low-dose and standard-dose scans for these cases. Using linear regression, the correlation of the automatic AS with the standard-dose manual scores was 0.86; with the low-dose manual scores the correlation was 0.91. Standard risk categories were also computed. The automated method risk category agreed with manual markings of gated scans for 24 cases while 15 cases were 1 category off. For low-dose scans, the automatic method agreed with 33 cases while 7 cases were 1 category off.

To evaluate commercial deep learning-based software for fully automated coronary artery calcium (CAC) scoring on non-electrocardiogram (ECG)-gated low-dose CT (LDCT) with different slice thicknesses compared with manual ECG-gated calcium-scoring CT (CSCT). This retrospective study included 567 patients who underwent both LDCT and CSCT. All LDCT images were reconstructed with a 2.5-mm slice thickness (LDCT2.5-mm), and 453 LDCT scans were reconstructed with a 1.0-mm slice thickness (LDCT1.0-mm). Automated CAC scoring was performed on CSCT (CSCTauto), LDCT1.0-mm, and LDCT2.5-mm images. The reliability of CSCTauto, LDCT1.0-mm, and LDCT2.5-mm was compared with manual CSCT scoring (CSCTmanual) using intraclass correlation coefficients (ICCs) and Bland-Altman analysis. Agreement, in CAC severity category, was analyzed using weighted kappa statistics. Diagnostic performance at various Agatston score cutoffs was also calculated. CSCTauto, LDCT1.0-mm, and LDCT2.5-mm demonstrated excellent agreement with CSCTmanual (ICC [95% confidence interval, CI]: 1.000 [1.000, 1.000], 0.937 [0.917, 0.952], and 0.955 [0.946, 0.963], respectively). The mean difference with 95% limits of agreement was lower with LDCT1.0-mm than with LDCT2.5-mm (19.94 [95% CI, -244.0, 283.9] vs. 45.26 [-248.2, 338.7]). Regarding CAC severity, LDCT1.0-mm achieved almost perfect agreement, and LDCT2.5-mm achieved substantial agreement (kappa [95% CI]: 0.809 [0.776, 0.838], 0.776 [0.740, 0.809], respectively). Diagnostic performance for detecting Agatston score ≥ 400 was also higher with LDCT1.0-mm than with LDCT2.5-mm (F1 score, 0.929 vs. 0.855). Fully automated CAC-scoring software with both CSCT and LDCT yielded excellent reliability and agreement with CSCTmanual. LDCT1.0-mm yielded more accurate Agatston scoring than LDCT2.5-mm using fully automated commercial software. • Total Agatston scores and all vessels of CSCTauto, LDCT1.0-mm, and LDCT2.5-mm demonstrated excellent agreement with CSCTmanual (all ICC > 0.85). • The diagnostic performance for detecting all Agatston score cutoffs was better with LDCT1.0-mm than with LDCT2.5-mm. • This automated software yielded a lower degree of underestimation compared with methods described in previous studies, and the degree of underestimation was lower with LDCT1.0-mm than with LDCT2.5-mm.

Deep learning based denoising techniques have effectively reduced noise in low-dose computed tomography (LDCT) images. However, the need for pairs of LDCT and normal-dose CT (NDCT) training images is the main reason that makes the use of deep learning networks difficult. Therefore, we propose a method of generating synthesized LDCT images by adding synthesized quantum noise to the NDCT images. For this, we estimate the quantum noise power spectrum (NPS) of patient CT images by modeling anatomical NPS and quantum NPS, respectively. We use the quantum NPS as a noise filter and generate synthesized quantum noise by filtering random Gaussian noise. Using the proposed method, LDCT denoiser can be trained without using pairs of LDCT and NDCT images, and the trained denoiser effectively reduces noise in LDCT images. The proposed method facilitates data acquisition for training deep learning networks and reduces the risk of radiation exposure to patients.

The valuable structure features in full-dose computed tomography (FdCT) scans can be exploited as prior knowledge for low-dose CT (LdCT) imaging. However, lacking the capability to represent local characteristics of interested structures of the LdCT image adaptively may result in poor preservation of details/textures in LdCT image. This paper aims to explore a novel prior knowledge retrieval and representation paradigm, called adaptive prior features assisted restoration algorithm, for the purpose of better restoration of the low-dose lung CT images by capturing local features from FdCT scans adaptively. The innovation lies in the construction of an offline training database and the online patch-search scheme integrated with the principal components analysis (PCA). Specifically, the offline training database is composed of 3-D patch samples extracted from existing full-dose lung scans. For online patch-search, 3-D patches with structure similar to the noisy target patch are first selected from the database as the training samples. Then, PCA is applied on the training samples to retrieve their local prior principal features adaptively. By employing the principal features to decompose the noisy target patch and using an adaptive coefficient shrinkage technique for inverse transformation, the noise of the target patch can be efficiently removed and the detailed texture can be well preserved. The effectiveness of the proposed algorithm was validated by CT scans of patients with lung cancer. The results show that it can achieve a noticeable gain over some state-of-the-art methods in terms of noise suppression and details/textures preservation.

Compared to other traditional imaging exams, computed tomography (CT) is more efficient, where a digital geometry processing is used to generate a 3D image of an internal structure of an object, or patient, from a series of 2D images obtained during various rotations of the CT scan around the scanned object. Also taking in consideration traditional imaging exams such as MRI or ultrasound, for example, the CT technique uses higher radiation doses than these exams, providing high quality images. However, in order to prevent constant exposures to high radiation doses, low-dose computed tomography (LDCT) scans are often recommended. Nevertheless, the images acquired in LDCT scans are degraded by undesirable artifacts, known as noise, which affects negatively the image quality. In this study, a two-stage filter based on morphological operators and Block-Matching 3D (BM3D) is proposed to remove noise in low-dose dental CT images. The quantitative results obtained by our proposed method demonstrated superior performance when compared to several state of the art techniques. Also, our proposed method obtained better visual performance, removing the noise and preserving details more efficiently than the compared filters.

Low-dose computed tomography (LDCT) is a diagnostic imaging technique designed to minimize radiation exposure to the patient. However, this reduction in radiation may compromise computed tomography (CT) image quality, adversely impacting clinical diagnoses. Various advanced LDCT methods have emerged to mitigate this challenge, relying on well-matched LDCT and normal-dose CT (NDCT) image pairs for training. Nevertheless, these methods often face difficulties in distinguishing image details from nonuniformly distributed noise, limiting their denoising efficacy. Additionally, acquiring suitably paired datasets in the medical domain poses challenges, further constraining their applicability. Hence, the objective of this study was to develop an innovative denoising framework for LDCT images employing unpaired data. In this paper, we propose a LDCT denoising network (DNCNN) that alleviates the need for aligning LDCT and NDCT images. Our approach employs generative adversarial networks (GANs) to learn and model the noise present in LDCT images, establishing a mapping from the pseudo-LDCT to the actual NDCT domain without the need for paired CT images. Within the domain of weakly supervised methods, our proposed model exhibited superior objective metrics on the simulated dataset when compared to CycleGAN and selective kernel-based cycle-consistent GAN (SKFCycleGAN): the peak signal-to-noise ratio (PSNR) was 43.9441, the structural similarity index measure (SSIM) was 0.9660, and the visual information fidelity (VIF) was 0.7707. In the clinical dataset, we conducted a visual effect analysis by observing various tissues through different observation windows. Our proposed method achieved a no-reference structural sharpness (NRSS) value of 0.6171, which was closest to that of the NDCT images (NRSS =0.6049), demonstrating its superiority over other denoising techniques in preserving details, maintaining structural integrity, and enhancing edge contrast. Through extensive experiments on both simulated and clinical datasets, we demonstrated the superior efficacy of our proposed method in terms of denoising quality and quantity. Our method exhibits superiority over both supervised techniques, including block-matching and 3D filtering (BM3D), residual encoder-decoder convolutional neural network (RED-CNN), and Wasserstein generative adversarial network-VGG (WGAN-VGG), and over weakly supervised approaches, including CycleGAN and SKFCycleGAN.

One-stage pulmonary nodule detection using 3-D DCNN with feature fusion and attention mechanism in CT image

Convolutional neural networks (CNN) have been used for a wide variety of deep learning applications, especially in computer vision. For medical image processing, researchers have identified certain challenges associated with CNNs. These challenges encompass the generation of less informative features, limitations in capturing both high and low-frequency information within feature maps, and the computational cost incurred when enhancing receptive fields by deepening the network. Transformers have emerged as an approach aiming to address and overcome these specific limitations of CNNs in the context of medical image analysis. Preservation of all spatial details of medical images is necessary to ensure accurate patient diagnosis. Hence, this research introduced the use of a pure Vision Transformer (ViT) for a denoising artificial neural network for medical image processing specifically for low-dose computed tomography (LDCT) image denoising. The proposed model follows a U-Net framework that contains ViT modules with the integration of Noise2Neighbor (N2N) interpolation operation. Five different datasets containing LDCT and normal-dose CT (NDCT) image pairs were used to carry out this experiment. To test the efficacy of the proposed model, this experiment includes comparisons between the quantitative and visual results among CNN-based (BM3D, RED-CNN, DRL-E-MP), hybrid CNN-ViT-based (TED-Net), and the proposed pure ViT-based denoising model. The findings of this study showed that there is about 15-20% increase in SSIM and PSNR when using self-attention transformers than using the typical pure CNN. Visual results also showed improvements especially when it comes to showing fine structural details of CT images.

This study aimed to evaluate image quality, the detection rate of enlarged lymph nodes, and radiation dose exposure of ultralow-dose and low-dose abdominopelvic computed tomography (CT) in patients with lymphoma. Patients with lymphoma who underwent abdominopelvic CT using dual-source scanner were retrospectively recruited from a single center. CT images were obtained at 90 kVp dual-source mode reformatted in three data sets using the advanced modelled iterative reconstruction algorithm: 100% (standard-dose CT), 66.7% (low-dose CT), and 33.3% (ultralow-dose CT). Two radiologists analyzed subjective image quality and detection of abdominal enlarged lymph nodes on ultralow-dose, low-dose, and standard-dose CT blindly and independently. The results were compared with reference standards. Three readers (two radiologists and one hematologist) reviewed overall image quality and spleen size. In total, 128 consecutive CT scans (77 complete response, 44 partial response, 6 progressive disease, and 1 initial evaluation) from 86 patients (64 B-cell lymphoma, 14 T/NK-cell lymphoma, and 8 Hodgkin’s lymphoma cases) were assessed. The enlarged lymph node-based detection rates for two readers were 97.0% (96/99) and 94.0% (93/99) on standard-dose CT, 97.0% (96/99) and 94.0% (93/99) on low-dose CT, and 94.0% (93/99) and 89.9% (89/99) on ultralow-dose CT. Overall image quality was 3.8 ± 0.5, 3.9 ± 0.5, and 4.1 ± 0.5 on ultralow-dose CT; 4.7 ± 0.4, 4.6 ± 0.5, and 4.8 ± 0.3 on low-dose CT; and 4.8 ± 0.4, 4.7 ± 0.4, and 4.9 ± 0.2 on standard-dose CT, according to two radiologists and one hematologist, respectively. Intraclass correlation coefficients of spleen size were 0.90 (95% confidence interval [CI], 0.87–0.93), 0.91 (95% CI, 0.88–0.93), and 0.91 (95% CI, 0.88–0.93) on ultralow-dose, low-dose, and standard-dose CT, respectively. Mean effective radiation doses of standard-dose, low-dose, and ultralow-dose CT were 5.7 ±1.8 mSv, 3.8 ± 1.2 mSv, and 1.9 ± 0.6 mSv, respectively. Our findings suggest that ultralow-dose and low-dose CT, even with radiation doses reduced by 66.7% and 33.3%, respectively, maintained adequate image quality. These imaging modalities may be employed for follow-up lymphoma evaluation in consideration of the long surveillance periods.

The resurgence of deep neural networks has created an alternative pathway for low-dose computed tomography denoising by learning a nonlinear transformation function between low-dose CT (LDCT) and normal-dose CT (NDCT) image pairs. However, those paired LDCT and NDCT images are rarely available in the clinical environment, making deep neural network deployment infeasible. This study proposes a novel method for self-supervised low-dose CT denoising to alleviate the requirement of paired LDCT and NDCT images. Specifically, we have trained an invertible neural network to minimize the pixel-based mean square distance between a noisy slice and the average of its two immediate adjacent noisy slices. We have shown the aforementioned is similar to training a neural network to minimize the distance between clean NDCT and noisy LDCT image pairs. Again, during the reverse mapping of the invertible network, the output image is mapped to the original input image, similar to cycle consistency loss. Finally, the trained invertible network’s forward mapping is used for denoising LDCT images. Extensive experiments on two publicly available datasets showed that our method performs favourably against other existing unsupervised methods.

Positron emission tomography (PET) and computed tomography (CT) play a vital role in tumor-related medical diagnosis, assessment, and treatment planning. However, full-dose PET and CT pose the risk of excessive radiation exposure to patients, whereas low-dose images compromise image quality, impacting subsequent tumor recognition and diseasediagnosis. To solve such problems, we propose a Noise-Assisted Hybrid Attention Network (NAHANet) to reconstruct full-dose PET and CT images from low-dose PET (LDPET) and CT (LDCT) images to reduce patient radiation risks while ensuring the performance of subsequent tumor recognition. NAHANet contains two branches: the noise feature prediction branch (NFPB) and the cascaded reconstruction branch. Among them, NFPB providing noise features for the cascade reconstruction branch. The cascaded reconstruction branch comprises a shallow feature extraction module and a reconstruction module which contains a series of cascaded noise feature fusion blocks (NFFBs). Among these, the NFFB fuses the features extracted from low-dose images with the noise features obtained by NFPB to improve the feature extraction capability. To validate the effectiveness of the NAHANet method, we performed experiments using two public available datasets: the Ultra-low Dose PET Imaging Challenge dataset and Low Dose CT Grand Challengedataset. As a result, the proposed NAHANet achieved higher performance on common indicators. For example, on the CT dataset, the PSNR and SSIM indicators were improved by 4.1dB and 0.06 respectively, and the rMSE indicator was reduced by 5.46 compared with the LDCT; on the PET dataset, the PSNR and SSIM was improved by 3.37dB and 0.02, and the rMSE was reduced by 9.04 compared with theLDPET. This paper proposes a transformer-based denoising algorithm, which utilizes hybrid attention to extract high-level features of low dose images and fuses noise features to optimize the denoising performance of the network, achieving good performance improvements on low-dose CT and PETdatasets.

Evaluation for Osteoporosis Using Low-Dose CT Imaging of the Chest Obtained for Lung Cancer Screening: A Retrospective Study of 1,336 Patients.

Low-dose computed tomography (CT) is of great potential advantage for disease diagnosis. Usually, paired training datasets are difficult to obtain in clinical routine, which catalyzes the development of unsupervised learning techniques to improve the low-dose CT imaging. Recently, most existing unsupervised learning approaches for low-dose CT imaging were developed in the image domain, and only a few approaches have been developed in the sinogram domain, which is a challenging task. In this paper, we propose a dedicated unpaired learning technique for low-dose CT sinogram restoration with a novel data-dependent noise-generative model. The general idea is to construct a paired pseudo normal-/low-dose sinogram dataset based on the existing unpaired normal-/low-dose sinogram dataset, after which a sinogram restoration network can be obtained by training on the paired pseudo normal-/low-dose sinogram dataset. However, the difficulty of the presented idea lies in the construction of the pseudo low-dose sinogram generative network, due to the complexity of the texture feature and noise property in the sinogram domain. To address this issue, we construct an appropriative generative network architecture based on a reasonable noise-generative model in the sinogram domain, which can be used to produce pseudo low-dose sinogram data within an adversarial learning framework. To validate the proposed technique, a clinical dataset was adopted. Experimental results demonstrate that the proposed method can produce promising pseudo low-dose sinogram data, which is sufficient to train an effective sinogram restoration network. Both quantitative and qualitative measurements show that the proposed method can obtained promising low-dose CT imaging performance.