Low-Dose CT Image Denoising Using a Generative Adversarial Network With Wasserstein Distance and Perceptual Loss

Zero-Shot Medical Image Translation via Frequency-Guided Diffusion Models

Recently, the diffusion model has emerged as a superior generative model that can produce high quality and realistic images. However, for medical image translation, the existing diffusion models are deficient in accurately retaining structural information since the structure details of source domain images are lost during the forward diffusion process and cannot be fully recovered through learned reverse diffusion, while the integrity of anatomical structures is extremely important in medical images. For instance, errors in image translation may distort, shift, or even remove structures and tumors, leading to incorrect diagnosis and inadequate treatments. Training and conditioning diffusion models using paired source and target images with matching anatomy can help. However, such paired data are very difficult and costly to obtain, and may also reduce the robustness of the developed model to out-of-distribution testing data. We propose a frequency-guided diffusion model (FGDM) that employs frequency-domain filters to guide the diffusion model for structure-preserving image translation. Based on its design, FGDM allows zero-shot learning, as it can be trained solely on the data from the target domain, and used directly for source-to-target domain translation without any exposure to the source-domain data during training. We evaluated it on three cone-beam CT (CBCT)-to-CT translation tasks for different anatomical sites, and a cross-institutional MR imaging translation task. FGDM outperformed the state-of-the-art methods (GAN-based, VAE-based, and diffusion-based) in metrics of Fréchet Inception Distance (FID), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM), showing its significant advantages in zero-shot medical image translation.

Deep learning-based algorithms for low-dose CT imaging: A review

The computed tomography (CT) technique is extensively employed as an imaging modality in clinical settings. The radiation dose of CT, however, is significantly high, thereby raising concerns regarding the potential radiation damage it may cause. The reduction of X-ray exposure dose in CT scanning may result in a significant decline in imaging quality, thereby elevating the risk of missed diagnosis and misdiagnosis. The reduction of CT radiation dose and acquisition of high-quality images to meet clinical diagnostic requirements have always been a critical research focus and challenge in the field of CT. Over the years, scholars have conducted extensive research on enhancing low-dose CT (LDCT) imaging algorithms, among which deep learning-based algorithms have demonstrated superior performance. In this review, we initially introduced the conventional algorithms for CT image reconstruction along with their respective advantages and disadvantages. Subsequently, we provided a detailed description of four aspects concerning the application of deep neural networks in LDCT imaging process: preprocessing in the projection domain, post-processing in the image domain, dual-domain processing imaging, and direct deep learning-based reconstruction (DLR). Furthermore, an analysis was conducted to evaluate the merits and demerits of each method. The commercial and clinical applications of the LDCT-DLR algorithm were also presented in an overview. Finally, we summarized the existing issues pertaining to LDCT-DLR and concluded the paper while outlining prospective trends for algorithmic advancement.

DeepmdQCT: A multitask network with domain invariant features and comprehensive attention mechanism for quantitative computer tomography diagnosis of osteoporosis

In the medical field, the application of machine learning technology in the automatic diagnosis and monitoring of osteoporosis often faces challenges related to domain adaptation in drug therapy research. The existing neural networks used for the diagnosis of osteoporosis may experience a decrease in model performance when applied to new data domains due to changes in radiation dose and equipment. To address this issue, in this study, we propose a new method for multi domain diagnostic and quantitative computed tomography (QCT) images, called DeepmdQCT. This method adopts a domain invariant feature strategy and integrates a comprehensive attention mechanism to guide the fusion of global and local features, effectively improving the diagnostic performance of multi domain CT images. We conducted experimental evaluations on a self-created OQCT dataset, and the results showed that for dose domain images, the average accuracy reached 91%, while for device domain images, the accuracy reached 90.5%. our method successfully estimated bone density values, with a fit of 0.95 to the gold standard. Our method not only achieved high accuracy in CT images in the dose and equipment fields, but also successfully estimated key bone density values, which is crucial for evaluating the effectiveness of osteoporosis drug treatment. In addition, we validated the effectiveness of our architecture in feature extraction using three publicly available datasets. We also encourage the application of the DeepmdQCT method to a wider range of medical image analysis fields to improve the performance of multi-domain images.

A Review of deep learning methods for denoising of medical low-dose CT images

To prevent patients from being exposed to excess of radiation in CT imaging, the most common solution is to decrease the radiation dose by reducing the X-ray, and thus the quality of the resulting low-dose CT images (LDCT) is degraded, as evidenced by more noise and streaking artifacts. Therefore, it is important to maintain high quality CT image while effectively reducing radiation dose. In recent years, with the rapid development of deep learning technology, deep learning-based LDCT denoising methods have become quite popular because of their data-driven and high-performance features to achieve excellent denoising results. However, to our knowledge, no relevant article has so far comprehensively introduced and reviewed advanced deep learning denoising methods such as Transformer structures in LDCT denoising tasks. Therefore, based on the literatures related to LDCT image denoising published from year 2016-2023, and in particular from 2020 to 2023, this study presents a systematic survey of current situation, and challenges and future research directions in LDCT image denoising field. Four types of denoising networks are classified according to the network structure: CNN-based, Encoder-Decoder-based, GAN-based, and Transformer-based denoising networks, and each type of denoising network is described and summarized from the perspectives of structural features and denoising performances. Representative deep-learning denoising methods for LDCT are experimentally compared and analyzed. The study results show that CNN-based denoising methods capture image details efficiently through multi-level convolution operation, demonstrating superior denoising effects and adaptivity. Encoder-decoder networks with MSE loss, achieve outstanding results in objective metrics. GANs based methods, employing innovative generators and discriminators, obtain denoised images that exhibit perceptually a closeness to NDCT. Transformer-based methods have potential for improving denoising performances due to their powerful capability in capturing global information. Challenges and opportunities for deep learning based LDCT denoising are analyzed, and future directions are also presented.

Proj2Proj: self-supervised low-dose CT reconstruction

In Computed Tomography (CT) imaging, one of the most serious concerns has always been ionizing radiation. Several approaches have been proposed to reduce the dose level without compromising the image quality. With the emergence of deep learning, thanks to the increasing availability of computational power and huge datasets, data-driven methods have recently received a lot of attention. Deep learning based methods have also been applied in various ways to address the low-dose CT reconstruction problem. However, the success of these methods largely depends on the availability of labeled data. On the other hand, recent studies showed that training can be done successfully without the need for labeled datasets. In this study, a training scheme was defined to use low-dose projections as their own training targets. The self-supervision principle was applied in the projection domain. The parameters of a denoiser neural network were optimized through self-supervised training. It was shown that our method outperformed both traditional and compressed sensing-based iterative methods, and deep learning based unsupervised methods, in the reconstruction of analytic CT phantoms and human CT images in low-dose CT imaging. Our method's reconstruction quality is also comparable to a well-known supervised method.

Semi-supervised low-dose SPECT restoration using sinogram inner-structure aware graph neural network

Objective.To mitigate the potential radiation risk, low-dose single photon emission computed tomography (SPECT) is of increasing interest. Numerous deep learning-based methods have been developed to perform low-dose imaging while maintaining image quality. However, most existing methods seldom explore the unique inner-structure inherent within sinograms. In addition, traditional supervised learning methods require large-scale labeled data, where the normal-dose data serves as annotation and is intractable to acquire in low-dose imaging. In this study, we aim to develop a novel sinogram inner-structure-aware semi-supervised framework for the task of low-dose SPECT sinogram restoration.Approach.The proposed framework retains the strengths of UNet, meanwhile introducing a sinogram-structure-based non-local neighbors graph neural network (SSN-GNN) module and a window-based K-nearest neighbors GNN (W-KNN-GNN) module to effectively exploit the inherent inner-structure within SPECT sinograms. Moreover, the proposed framework employs the mean teacher semi-supervised learning approach to leverage the information available in abundant unlabeled low-dose sinograms.Main results.The datasets exploited in this study were acquired from the (Extended Cardiac-Torso) XCAT anthropomorphic digital phantoms, which provide realistic images for imaging research of various modalities. Quantitative as well as qualitative results demonstrate that the proposed framework achieves superior performance compared to several state-of-the-art reconstruction methods. To further validate the effectiveness of the proposed framework, ablation and robustness experiments were also performed. The experimental results show that each component of the proposed framework effectively improves the model performance, and the framework exhibits superior robustness with respect to various noise levels. Besides, the proposed semi-supervised paradigm showcases the efficacy of incorporating supplementary unlabeled low-dose sinograms.Significance.The proposed framework improves the quality of low-dose SPECT reconstructed images by utilizing sinogram inner-structure and incorporating supplementary unlabeled data, which provides an important tool for dose reduction without sacrificing the image quality.

[A low- dose CT reconstruction algorithm across different scanners based on federated feature learning]

OBJECTIVE

To propose a low-dose CT reconstruction algorithm across different scanners based on federated feature learning (FedCT) to improve the generalization of deep learning models for multiple CT scanners and protect data privacy.

METHODS

In the proposed FedCT framework, each client is assigned an inverse Radon transform-based reconstruction model to serve as a local network model that participates in federated learning. A projection- domain specific learning strategy is adopted to preserve the geometry specificity in the local projection domain. Federated feature learning is introduced in the model, which utilizes conditional parameters to mark the local data and feed the conditional parameters into the network for encoding to enhance the generalization of the model in the image domain.

RESULTS

In the cross-client, multi-scanner, and multi-protocol low-dose CT reconstruction experiments, FedCT achieved the highest PSNR (+2.8048, +2.7301, and +2.7263 compared to the second best federated learning method), the highest SSIM (+0.0009, +0.0165, and +0.0131 in the same comparison), and the lowest RMSE (- 0.6687, - 1.5956, and - 0.9962). In the ablation experiment, compared with the general federated learning strategy, the model with projection-specific learning strategy showed an average improvement by 1.18 on Q1 of the PSNR and an average decrease by 1.36 on Q3 of the RMSE on the test set. The introduction of federated feature learning in FedCT further improved the Q1 of the PSNR on the test set by 3.56 and reduced the Q3 of the RMSE by 1.80.

CONCLUSION

FedCT provides an effective solution for collaborative construction of CT reconstruction models, which can enhance model generalization and further improve the reconstruction performance on global data while protecting data privacy.

Unpaired intra-operative OCT (iOCT) video super-resolution with contrastive learning

Regenerative therapies show promise in reversing sight loss caused by degenerative eye diseases. Their precise subretinal delivery can be facilitated by robotic systems alongside with Intra-operative Optical Coherence Tomography (iOCT). However, iOCT's real-time retinal layer information is compromised by inferior image quality. To address this limitation, we introduce an unpaired video super-resolution methodology for iOCT quality enhancement. A recurrent network is proposed to leverage temporal information from iOCT sequences, and spatial information from pre-operatively acquired OCT images. Additionally, a patchwise contrastive loss enables unpaired super-resolution. Extensive quantitative analysis demonstrates that our approach outperforms existing state-of-the-art iOCT super-resolution models. Furthermore, ablation studies showcase the importance of temporal aggregation and contrastive loss in elevating iOCT quality. A qualitative study involving expert clinicians also confirms this improvement. The comprehensive evaluation demonstrates our method's potential to enhance the iOCT image quality, thereby facilitating successful guidance for regenerative therapies.

CoreDiff: Contextual Error-Modulated Generalized Diffusion Model for Low-Dose CT Denoising and Generalization

Low-dose computed tomography (CT) images suffer from noise and artifacts due to photon starvation and electronic noise. Recently, some works have attempted to use diffusion models to address the over-smoothness and training instability encountered by previous deep-learning-based denoising models. However, diffusion models suffer from long inference time due to a large number of sampling steps involved. Very recently, cold diffusion model generalizes classical diffusion models and has greater flexibility. Inspired by cold diffusion, this paper presents a novel COntextual eRror-modulated gEneralized Diffusion model for low-dose CT (LDCT) denoising, termed CoreDiff. First, CoreDiff utilizes LDCT images to displace the random Gaussian noise and employs a novel mean-preserving degradation operator to mimic the physical process of CT degradation, significantly reducing sampling steps thanks to the informative LDCT images as the starting point of the sampling process. Second, to alleviate the error accumulation problem caused by the imperfect restoration operator in the sampling process, we propose a novel ContextuaL Error-modulAted Restoration Network (CLEAR-Net), which can leverage contextual information to constrain the sampling process from structural distortion and modulate time step embedding features for better alignment with the input at the next time step. Third, to rapidly generalize the trained model to a new, unseen dose level with as few resources as possible, we devise a one-shot learning framework to make CoreDiff generalize faster and better using only one single LDCT image (un)paired with normal-dose CT (NDCT). Extensive experimental results on four datasets demonstrate that our CoreDiff outperforms competing methods in denoising and generalization performance, with clinically acceptable inference time. Source code is made available at https://github.com/qgao21/CoreDiff.

Systematic Review on Learning-based Spectral CT

Spectral computed tomography (CT) has recently emerged as an advanced version of medical CT and significantly improves conventional (single-energy) CT. Spectral CT has two main forms: dual-energy computed tomography (DECT) and photon-counting computed tomography (PCCT), which offer image improvement, material decomposition, and feature quantification relative to conventional CT. However, the inherent challenges of spectral CT, evidenced by data and image artifacts, remain a bottleneck for clinical applications. To address these problems, machine learning techniques have been widely applied to spectral CT. In this review, we present the state-of-the-art data-driven techniques for spectral CT.

Learning Spatio-Temporal Pulse Representation With Global-Local Interaction and Supervision for Remote Prediction of Heart Rate

Recent studies have demonstrated the benefit of extracting and fusing pulse signals from multi-scale region-of-interests (ROIs). However, these methods suffer from heavy computational load. This paper aims to effectively utilize multi-scale rPPG features with a more compact architecture. Inspired by recent research works exploring two-path architecture that leverages global and local information with bidirectional bridge in between. This paper designs a novel architecture Global-Local Interaction and Supervision Network (GLISNet), which uses a local path to learn representations in the original scale and a global path to learn representations in the other scale capturing multi-scale information. A light-weight rPPG signal generation block is attached to the output of each path that maps the pulse representation to the pulse output. A hybrid loss function is utilized enabling the local and global representations to learn directly from the training data. Extensive experiments are conducted on two publicly available datasets, and results demonstrate that GLISNet achieves superior performance in terms of signal-to-noise ratio (SNR), mean absolute error (MAE), and root mean squared error (RMSE). In terms of SNR, GLISNet has an increase of 4.41% compared with the second best algorithm PhysNet on PURE dataset. The MAE has a decrease of 13.16% compared with the second best algorithm DeeprPPG on UBFC-rPPG dataset. The RMSE has a decrease of 26.29% compared with the second best algorithm PhysNet on UBFC-rPPG dataset. Experiments on MIHR dataset demonstrates the robustness of GLISNet under low-light environment.

An unsupervised two-step training framework for low-dose computed tomography denoising

BACKGROUND

Although low-dose computed tomography (CT) imaging has been more widely adopted in clinical practice to reduce radiation exposure to patients, the reconstructed CT images tend to have more noise, which impedes accurate diagnosis. Recently, deep neural networks using convolutional neural networks to reduce noise in the reconstructed low-dose CT images have shown considerable improvement. However, they need a large number of paired normal- and low-dose CT images to fully train the network via supervised learning methods.

PURPOSE

To propose an unsupervised two-step training framework for image denoising that uses low-dose CT images of one dataset and unpaired high-dose CT images from another dataset.

METHODS

Our proposed framework trains the denoising network in two steps. In the first training step, we train the network using 3D volumes of CT images and predict the center CT slice from them. This pre-trained network is used in the second training step to train the denoising network and is combined with the memory-efficient denoising generative adversarial network (DenoisingGAN), which further enhances both objective and perceptual quality.

RESULTS

The experimental results on phantom and clinical datasets show superior performance over the existing traditional machine learning and self-supervised deep learning methods, and the results are comparable to the fully supervised learning methods.

CONCLUSIONS

We proposed a new unsupervised learning framework for low-dose CT denoising, convincingly improving noisy CT images from both objective and perceptual quality perspectives. Because our denoising framework does not require physics-based noise models or system-dependent assumptions, our proposed method can be easily reproduced; consequently, it can also be generally applicable to various CT scanners or dose levels.

CT image denoising methods for image quality improvement and radiation dose reduction

With the ever-increasing use of computed tomography (CT), concerns about its radiation dose have become a significant public issue. To address the need for radiation dose reduction, CT denoising methods have been widely investigated and applied in low-dose CT images. Numerous noise reduction algorithms have emerged, such as iterative reconstruction and most recently, deep learning (DL)-based approaches. Given the rapid advancements in Artificial Intelligence techniques, we recognize the need for a comprehensive review that emphasizes the most recently developed methods. Hence, we have performed a thorough analysis of existing literature to provide such a review. Beyond directly comparing the performance, we focus on pivotal aspects, including model training, validation, testing, generalizability, vulnerability, and evaluation methods. This review is expected to raise awareness of the various facets involved in CT image denoising and the specific challenges in developing DL-based models.

Self-supervised Self2Self denoising strategy for OCT speckle reduction with a single noisy image

Optical coherence tomography (OCT) inevitably suffers from the influence of speckles originating from multiple scattered photons owing to its low-coherence interferometry property. Although various deep learning schemes have been proposed for OCT despeckling, they typically suffer from the requirement for ground-truth images, which are difficult to collect in clinical practice. To alleviate the influences of speckles without requiring ground-truth images, this paper presents a self-supervised deep learning scheme, namely, Self2Self strategy (S2Snet), for OCT despeckling using a single noisy image. Specifically, in this study, the main deep learning architecture is the Self2Self network, with its partial convolution being updated with a gated convolution layer. Specifically, both the input images and their Bernoulli sampling instances are adopted as network input first, and then, a devised loss function is integrated into the network to remove the background noise. Finally, the denoised output is estimated using the average of multiple predicted outputs. Experiments with various OCT datasets are conducted to verify the effectiveness of the proposed S2Snet scheme. Results compared with those of the existing methods demonstrate that S2Snet not only outperforms those existing self-supervised deep learning methods but also achieves better performances than those non-deep learning ones in different cases. Specifically, S2Snet achieves an improvement of 3.41% and 2.37% for PSNR and SSIM, respectively, as compared to the original Self2Self network, while such improvements become 19.9% and 22.7% as compared with the well-known non-deep learning NWSR method.

PIDNET: Polar Transformation Based Implicit Disentanglement Network for Truncation Artifacts

The interior problem, a persistent ill-posed challenge in CT imaging, gives rise to truncation artifacts capable of distorting CT values, thereby significantly impacting clinical diagnoses. Traditional methods have long struggled to effectively solve this issue until the advent of supervised models built on deep neural networks. However, supervised models are constrained by the need for paired data, limiting their practical application. Therefore, we propose a simple and efficient unsupervised method based on the Cycle-GAN framework. Introducing an implicit disentanglement strategy, we aim to separate truncation artifacts from content information. The separated artifact features serve as complementary constraints and the source of generating simulated paired data to enhance the training of the sub-network dedicated to removing truncation artifacts. Additionally, we incorporate polar transformation and an innovative constraint tailored specifically for truncation artifact features, further contributing to the effectiveness of our approach. Experiments conducted on multiple datasets demonstrate that our unsupervised network outperforms the traditional Cycle-GAN model significantly. When compared to state-of-the-art supervised models trained on paired datasets, our model achieves comparable visual results and closely aligns with quantitative evaluation metrics.

Deep Filtered Back Projection for CT Reconstruction

Filtered back projection (FBP) is a classic analytical algorithm for computed tomography (CT) reconstruction, with high computational efficiency. However, images reconstructed by FBP often suffer from excessive noise and artifacts. The original FBP algorithm uses a window function to smooth signals and a linear interpolation to estimate projection values at un-sampled locations. In this study, we propose a novel framework named DeepFBP in which an optimized filter and an optimized nonlinear interpolation operator are learned with neural networks. Specifically, the learned filter can be considered as the product of an optimized window function and the ramp filter, and the learned interpolation can be considered as an optimized way to utilize projection information of nearby locations through nonlinear combination. The proposed method remains the high computational efficiency of the original FBP and achieves much better reconstruction quality at different noise levels. It also outperforms the TV-based statistical iterative algorithm, with computational time being reduced in an order of two, and state-of-the-art post-processing deep learning methods that have deeper and more complicated network structures.

Combined Filtering Method for Offshore Oil and Gas Platform Point Cloud Data Based on KNN_PCF and Hy_WHF and Its Application in 3D Reconstruction

With the increasing scale of deep-sea oil exploration and drilling platforms, the assessment, maintenance, and optimization of marine structures have become crucial. Traditional detection and manual measurement methods are inadequate for meeting these demands, but three-dimensional laser scanning technology offers a promising solution. However, the complexity of the marine environment, including waves and wind, often leads to problematic point cloud data characterized by noise points and redundancy. To address this challenge, this paper proposes a method that combines K-Nearest-Neighborhood filtering with a hyperbolic function-based weighted hybrid filtering. The experimental results demonstrate the exceptional performance of the algorithm in processing point cloud data from offshore oil and gas platforms. The method improves noise point filtering efficiency by approximately 11% and decreases the total error by 0.6 percentage points compared to existing technologies. Not only does this method accurately process anomalies in high-density areas-it also removes noise while preserving important details. Furthermore, the research method presented in this paper is particularly suited for processing large point cloud data in complex marine environments. It enhances data accuracy and optimizes the three-dimensional reconstruction of offshore oil and gas platforms, providing reliable dimensional information for land-based prefabrication of these platforms.

A denoising method based on deep learning for proton radiograph using energy resolved dose function

Objective.Proton radiograph has been broadly applied in proton radiotherapy which is affected by scattered protons which result in the lower spatial resolution of proton radiographs than that of x-ray images. Traditional image denoising method may lead to the change of water equivalent path length (WEPL) resulting in the lower WEPL measurement accuracy. In this study, we proposed a new denoising method of proton radiographs based on energy resolved dose function curves.Approach.Firstly, the corresponding relationship between the distortion of WEPL characteristic curve, and energy and proportion of scattered protons was established. Then, to improve the accuracy of proton radiographs, deep learning technique was used to remove scattered protons and correct deviated WEPL values. Experiments on a calibration phantom to prove the effectiveness and feasibility of this method were performed. In addition, an anthropomorphic head phantom was selected to demonstrate the clinical relevance of this technology and the denoising effect was analyzed.Main results.The curves of WEPL profiles of proton radiographs became smoother and deviated WEPL values were corrected. For the calibration phantom proton radiograph, the average absolute error of WEPL values decreased from 2.23 to 1.72, the mean percentage difference of all materials of relative stopping power decreased from 1.24 to 0.39, and the average relative WEPL corrected due to the denoising process was 1.06%. In addition, WEPL values correcting were also observed on the proton radiograph for anthropomorphic head phantom due to this denoising process.Significance.The experiments showed that this new method was effective for proton radiograph denoising and had greater advantages than end-to-end image denoising methods, laying the foundation for the implementation of precise proton radiotherapy.

DiffusionCT: Latent Diffusion Model for CT Image Standardization

Computed tomography (CT) is one of the modalities for effective lung cancer screening, diagnosis, treatment, and prognosis. The features extracted from CT images are now used to quantify spatial and temporal variations in tumors. However, CT images obtained from various scanners with customized acquisition protocols may introduce considerable variations in texture features, even for the same patient. This presents a fundamental challenge to downstream studies that require consistent and reliable feature analysis. Existing CT image harmonization models rely on GAN-based supervised or semi-supervised learning, with limited performance. This work addresses the issue of CT image harmonization using a new diffusion-based model, named DiffusionCT, to standardize CT images acquired from different vendors and protocols. DiffusionCT operates in the latent space by mapping a latent non-standard distribution into a standard one. DiffusionCT incorporates a U-Net-based encoder-decoder, augmented by a diffusion model integrated into the bottleneck part. The model is designed in two training phases. The encoder-decoder is first trained, without embedding the diffusion model, to learn the latent representation of the input data. The latent diffusion model is then trained in the next training phase while fixing the encoder-decoder. Finally, the decoder synthesizes a standardized image with the transformed latent representation. The experimental results demonstrate a significant improvement in the performance of the standardization task using DiffusionCT.

M3S-Net: multi-modality multi-branch multi-self-attention network with structure-promoting loss for low-dose PET/CT enhancement

Objective.PET (Positron Emission Tomography) inherently involves radiotracer injections and long scanning time, which raises concerns about the risk of radiation exposure and patient comfort. Reductions in radiotracer dosage and acquisition time can lower the potential risk and improve patient comfort, respectively, but both will also reduce photon counts and hence degrade the image quality. Therefore, it is of interest to improve the quality of low-dose PET images.Approach.A supervised multi-modality deep learning model, named M3S-Net, was proposed to generate standard-dose PET images (60 s per bed position) from low-dose ones (10 s per bed position) and the corresponding CT images. Specifically, we designed a multi-branch convolutional neural network with multi-self-attention mechanisms, which first extracted features from PET and CT images in two separate branches and then fused the features to generate the final generated PET images. Moreover, a novel multi-modality structure-promoting term was proposed in the loss function to learn the anatomical information contained in CT images.Main results.We conducted extensive numerical experiments on real clinical data collected from local hospitals. Compared with state-of-the-art methods, the proposed M3S-Net not only achieved higher objective metrics and better generated tumors, but also performed better in preserving edges and suppressing noise and artifacts.Significance.The experimental results of quantitative metrics and qualitative displays demonstrate that the proposed M3S-Net can generate high-quality PET images from low-dose ones, which are competable to standard-dose PET images. This is valuable in reducing PET acquisition time and has potential applications in dynamic PET imaging.

Deep learning-based projection synthesis for low-dose cone-beam computed tomography imaging in image-guided radiotherapy

Background

The imaging dose of cone-beam computed tomography (CBCT) in image-guided radiotherapy (IGRT) poses adverse effects on patient health. To improve the quality of sparse-view low-dose CBCT images, a projection synthesis convolutional neural network (SynCNN) model is proposed.

Methods

Included in this retrospective, single-center study were 223 patients diagnosed with brain tumours from Beijing Cancer Hospital. The proposed SynCNN model estimated two pairs of orthogonally direction-separable spatial kernels to synthesize the missing projection in between the input neighboring sparse-view projections via local convolution operations. The SynCNN model was trained on 150 real patients to learn patterns for inter-view projection synthesis. CBCT data from 30 real patients were used to validate the SynCNN, while data from a phantom and 43 real patients were used to test the SynCNN externally. Sparse-view projection datasets with 1/2, 1/4, and 1/8 of the original sampling rate were simulated, and the corresponding full-view projection datasets were restored using the SynCNN model. The tomographic images were then reconstructed with the Feldkamp-Davis-Kress algorithm. The root-mean-square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) metrics were measured in both the projection and image domains. Five experts were invited to grade the image quality blindly for 40 randomly selected evaluation groups with a four-level rubric, where a score greater than or equal to 2 was considered acceptable image quality. The running time of the SynCNN model was recorded. The SynCNN model was directly compared with the three other methods on 1/4 sparse-view reconstructions.

Results

The phantom and patient studies showed that the missing projections were accurately synthesized. In the image domain, for the phantom study, compared with images reconstructed from sparse-view projections, images with SynCNN synthesis exhibited significantly improved qualities with decreased values in RMSE and increased values in PSNR and SSIM. For the patient study, between the results with and without the SynCNN synthesis, the averaged RMSE decreased by 3.4×10-4, 10.3×10-4, and 21.7×10-4, the averaged PSNR increased by 3.4, 6.6, and 9.4 dB, and the averaged SSIM increased by 5.2×10-2, 18.9×10-2 and 33.9×10-2, for the 1/2, 1/4, and 1/8 sparse-view reconstructions, respectively. In expert subjective evaluation, both the median scores and acceptance rates of the images with SynCNN synthesis were higher than those reconstructed from sparse-view projections. It took the model less than 0.01 s to synthesize an inter-view projection. Compared with the three other methods, the SynCNN model obtained the best scores in terms of the three metrics in both domains.

Conclusions

The proposed SynCNN model effectively improves the quality of sparse-view CBCT images at a low time cost. With the SynCNN model, the CBCT imaging dose in IGRT could be reduced potentially.

A two-stage deep-learning framework for CT denoising based on a clinically structure-unaligned paired data set

Background

In low-dose computed tomography (LDCT) lung cancer screening, soft tissue is hardly appreciable due to high noise levels. While deep learning-based LDCT denoising methods have shown promise, they typically rely on structurally aligned synthesized paired data, which lack consideration of the clinical reality that there are no aligned LDCT and normal-dose CT (NDCT) images available. This study introduces an LDCT denoising method using clinically structure-unaligned but paired data sets (LDCT and NDCT scans from the same patients) to improve lesion detection during LDCT lung cancer screening.

Methods

A cohort of 64 patients undergoing both LDCT and NDCT was randomly divided into training (n=46) and testing (n=18) sets. A two-stage training approach was adopted. First, Gaussian noise was added to NDCT data to create simulated LDCT data for generator training. Then, the model was trained on a clinically structure-unaligned paired data set using a Wasserstein generative adversarial network (WGAN) framework with the initial generator weights obtained during the first stage of training. An attention mechanism was also incorporated into the network.

Results

Validated on a clinical CT data set, our proposed method outperformed other available methods [CycleGAN, Pixel2Pixel, block-matching and three-dimensional filtering (BM3D)] in noise removal and detail retention tasks in terms of the peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and root mean square error (RMSE) metrics. Compared with the results produced by BM3D, our method yielded an average improvement of approximately 7% in terms of the three evaluation indicators. The probability density profile of the denoised CT output produced using our method best fit the reference NDCT scan. Additionally, our two-stage model outperformed the one-stage WGAN-based model in both objective and subjective evaluations, further demonstrating the higher effectiveness of our two-stage training approach.

Conclusions

The proposed method performed the best in removing noise from LDCT scans and exhibited good detail retention, which could potentially enhance the lesion detection and characterization effects obtained for soft tissues in the scanning scope of LDCT lung cancer screening.

GAN-based medical image small region forgery detection via a two-stage cascade framework

Using generative adversarial network (GAN) Goodfellow et al. (2014) for data enhancement of medical images is significantly helpful for many computer-aided diagnosis (CAD) tasks. A new GAN-based automated tampering attack, like CT-GAN Mirsky et al. (2019), has emerged. It can inject or remove lung cancer lesions to CT scans. Because the tampering region may even account for less than 1% of the original image, even state-of-the-art methods are challenging to detect the traces of such tampering. This paper proposes a two-stage cascade framework to detect GAN-based medical image small region forgery like CT-GAN. In the local detection stage, we train the detector network with small sub-images so that interference information in authentic regions will not affect the detector. We use depthwise separable convolution and residual networks to prevent the detector from over-fitting and enhance the ability to find forged regions through the attention mechanism. The detection results of all sub-images in the same image will be combined into a heatmap. In the global classification stage, using gray-level co-occurrence matrix (GLCM) can better extract features of the heatmap. Because the shape and size of the tampered region are uncertain, we use hyperplanes in an infinite-dimensional space for classification. Our method can classify whether a CT image has been tampered and locate the tampered position. Sufficient experiments show that our method can achieve excellent performance than the state-of-the-art detection methods.

Decomposition iteration strategy for low-dose CT denoising

In the medical field, computed tomography (CT) is a commonly used examination method, but the radiation generated increases the risk of illness in patients. Therefore, low-dose scanning schemes have attracted attention, in which noise reduction is essential. We propose a purposeful and interpretable decomposition iterative network (DISN) for low-dose CT denoising. This method aims to make the network design interpretable and improve the fidelity of details, rather than blindly designing or using deep CNN architecture. The experiment is trained and tested on multiple data sets. The results show that the DISN method can restore the low-dose CT image structure and improve the diagnostic performance when the image details are limited. Compared with other algorithms, DISN has better quantitative and visual performance, and has potential clinical application prospects.

Deep-silicon photon-counting x-ray projection denoising through reinforcement learning

BACKGROUND

In recent years, deep reinforcement learning (RL) has been applied to various medical tasks and produced encouraging results.

OBJECTIVE

In this paper, we demonstrate the feasibility of deep RL for denoising simulated deep-silicon photon-counting CT (PCCT) data in both full and interior scan modes. PCCT offers higher spatial and spectral resolution than conventional CT, requiring advanced denoising methods to suppress noise increase.

METHODS

In this work, we apply a dueling double deep Q network (DDDQN) to denoise PCCT data for maximum contrast-to-noise ratio (CNR) and a multi-agent approach to handle data non-stationarity.

RESULTS

Using our method, we obtained significant image quality improvement for single-channel scans and consistent improvement for all three channels of multichannel scans. For the single-channel interior scans, the PSNR (dB) and SSIM increased from 33.4078 and 0.9165 to 37.4167 and 0.9790 respectively. For the multichannel interior scans, the channel-wise PSNR (dB) increased from 31.2348, 30.7114, and 30.4667 to 31.6182, 30.9783, and 30.8427 respectively. Similarly, the SSIM improved from 0.9415, 0.9445, and 0.9336 to 0.9504, 0.9493, and 0.0326 respectively.

CONCLUSIONS

Our results show that the RL approach improves image quality effectively, efficiently, and consistently across multiple spectral channels and has great potential in clinical applications.

Synthetization of high-dose images using low-dose CT scans

BACKGROUND

Radiation dose reduction has been the focus of many research activities in x-ray CT. Various approaches were taken to minimize the dose to patients, ranging from the optimization of clinical protocols, refinement of the scanner hardware design, and development of advanced reconstruction algorithms. Although significant progress has been made, more advancements in this area are needed to minimize the radiation risks to patients.

PURPOSE

Reconstruction algorithm-based dose reduction approaches focus mainly on the suppression of noise in the reconstructed images while preserving detailed anatomical structures. Such an approach effectively produces synthesized high-dose images (SHD) from the data acquired with low-dose scans. A representative example is the model-based iterative reconstruction (MBIR). Despite its widespread deployment, its full adoption in a clinical environment is often limited by an undesirable image texture. Recent studies have shown that deep learning image reconstruction (DLIR) can overcome this shortcoming. However, the limited availability of high-quality clinical images for training and validation is often the bottleneck for its development. In this paper, we propose a novel approach to generate SHD with existing low-dose clinical datasets that overcomes both the noise texture issue and the data availability issue.

METHODS

Our approach is based on the observation that noise in the image can be effectively reduced by performing image processing orthogonal to the imaging plane. This process essentially creates an equivalent thick-slice image (TSI), and the characteristics of TSI depend on the nature of the image processing. An advantage of this approach is its potential to reduce impact on the noise texture. The resulting image, however, is likely corrupted by the anatomical structural degradation due to partial volume effects. Careful examination has shown that the differential signal between the original and the processed image contains sufficient information to identify regions where anatomical structures are modified. The differential signal, unfortunately, contains significant noise and has to be removed. The noise removal can be accomplished by performing iterative noise reduction to preserve structural information. The processed differential signal is subsequently subtracted from TSI to arrive at SHD.

RESULTS

The algorithm was evaluated extensively with phantom and clinical datasets. For better visual inspection, difference images between the original and SHD were generated and carefully examined. Negligible residual structure could be observed. In addition to the qualitative inspection, quantitative analyses were performed on clinical images in terms of the CT number consistency and the noise reduction characteristics. Results indicate that no CT number bias is introduced by the proposed algorithm. In addition, noise reduction capability is consistent across different patient anatomical regions. Further, simulated water phantom scans were utilized in the generation of the noise power spectrum (NPS) to demonstrate the preservation of the noise-texture.

CONCLUSIONS

We present a method to generate SHD datasets from regularly acquired low-dose CT scans. Images produced with the proposed approach exhibit excellent noise-reduction with the desired noise-texture. Extensive clinical and phantom studies have demonstrated the efficacy and robustness of our approach. Potential limitations of the current implementation are discussed and further research topics are outlined.

Multiobjective optimization guided by image quality index for limited-angle CT image reconstruction

BACKGROUND

Due to the incomplete projection data collected by limited-angle computed tomography (CT), severe artifacts are present in the reconstructed image. Classical regularization methods such as total variation (TV) minimization, ℓ0 minimization, are unable to suppress artifacts at the edges perfectly. Most existing regularization methods are single-objective optimization approaches, stemming from scalarization methods for multiobjective optimization problems (MOP).

OBJECTIVE

To further suppress the artifacts and effectively preserve the edge structures of the reconstructed image.

METHOD

This study presents a multiobjective optimization model incorporates both data fidelity term and ℓ0-norm of the image gradient as objective functions. It employs an iterative approach different from traditional scalarization methods, using the maximization of structural similarity (SSIM) values to guide optimization rather than minimizing the objective function.The iterative method involves two steps, firstly, simultaneous algebraic reconstruction technique (SART) optimizes the data fidelity term using SSIM and the Simulated Annealing (SA) algorithm for guidance. The degradation solution is accepted in the form of probability, and guided image filtering (GIF) is introduced to further preserve the image edge when the degradation solution is rejected. Secondly, the result from the first step is integrated into the second objective function as a constraint, we use ℓ0 minimization to optimize ℓ0-norm of the image gradient, and the SSIM, SA algorithm and GIF are introduced to guide optimization process by improving SSIM value like the first step.

RESULTS

With visual inspection, the peak signal-to-noise ratio (PSNR), root mean square error (RMSE), and SSIM values indicate that our approach outperforms other traditional methods.

CONCLUSIONS

The experiments demonstrate the effectiveness of our method and its superiority over other classical methods in artifact suppression and edge detail restoration.

APNet: Adaptive projection network for medical image denoising

BACKGROUND

In clinical medicine, low-dose radiographic image noise reduces the quality of the detected image features and may have a negative impact on disease diagnosis.

OBJECTIVE

In this study, Adaptive Projection Network (APNet) is proposed to reduce noise from low-dose medical images.

METHODS

APNet is developed based on an architecture of the U-shaped network to capture multi-scale data and achieve end-to-end image denoising. To adaptively calibrate important features during information transmission, a residual block of the dual attention method throughout the encoding and decoding phases is integrated. A non-local attention module to separate the noise and texture of the image details by using image adaptive projection during the feature fusion.

RESULTS

To verify the effectiveness of APNet, experiments on lung CT images with synthetic noise are performed, and the results demonstrate that the proposed approach outperforms recent methods in both quantitative index and visual quality. In addition, the denoising experiment on the dental CT image is also carried out and it verifies that the network has a certain generalization.

CONCLUSIONS

The proposed APNet is an effective method that can reduce image noise and preserve the required image details in low-dose radiographic images.

Learnable PM diffusion coefficients and reformative coordinate attention network for low dose CT denoising

Objective.Various deep learning methods have recently been used for low dose CT (LDCT) denoising. Aggressive denoising may destroy the edge and fine anatomical structures of CT images. Therefore a key issue in LDCT denoising tasks is the difficulty of balancing noise/artifact suppression and edge/structure preservation.Approach.We proposed an LDCT denoising network based on the encoder-decoder structure, namely the Learnable PM diffusion coefficient and efficient attention network (PMA-Net). First, using the powerful feature modeling capability of partial differential equations, we constructed a multiple learnable edge module to generate precise edge information, incorporating the anisotropic image processing idea of Perona-Malik (PM) model into the neural network. Second, a multiscale reformative coordinate attention module was designed to extract multiscale information. Non-overlapping dilated convolution capturing abundant contextual content was combined with coordinate attention which could embed the spatial location information of important features into the channel attention map. Finally, we imposed additional constraints on the edge information using edge-enhanced multiscale perceptual loss to avoid structure loss and over-smoothing.Main results.Experiments are conducted on simulated and real datasets. The quantitative and qualitative results show that the proposed method has better performance in suppressing noise/artifacts and preserving edges/structures.Significance.This work proposes a novel edge feature extraction method that unfolds partial differential equation into neural networks, which contributes to the interpretability and clinical application value of neural network.

Patient-derived PixelPrint phantoms for evaluating clinical imaging performance of a deep learning CT reconstruction algorithm

Objective

Deep learning reconstruction (DLR) algorithms exhibit object-dependent resolution and noise performance. Thus, traditional geometric CT phantoms cannot fully capture the clinical imaging performance of DLR. This study uses a patient-derived 3D-printed PixelPrint lung phantom to evaluate a commercial DLR algorithm across a wide range of radiation dose levels.

Approach

The lung phantom used in this study is based on a patient chest CT scan containing ground glass opacities and was fabricated using PixelPrint 3D-printing technology. The phantom was placed inside two different sized extension rings to mimic a small and medium sized patient and was scanned on a conventional CT scanner at exposures between 0.5 and 20 mGy. Each scan was reconstructed using filtered back projection (FBP), iterative reconstruction, and DLR at five levels of denoising. Image noise, contrast to noise ratio (CNR), root mean squared error (RMSE), structural similarity index (SSIM), and multi-scale SSIM (MS SSIM) were calculated for each image.

Main Results

DLR demonstrated superior performance compared to FBP and iterative reconstruction for all measured metrics in both phantom sizes, with better performance for more aggressive denoising levels. DLR was estimated to reduce dose by 25-83% in the small phantom and by 50-83% in the medium phantom without decreasing image quality for any of the metrics measured in this study. These dose reduction estimates are more conservative compared to the estimates obtained when only considering noise and CNR with a non-anatomical physics phantom.

Significance

DLR has the capability of producing diagnostic image quality at up to 83% lower radiation dose which can improve the clinical utility and viability of lower dose CT scans. Furthermore, the PixelPrint phantom used in this study offers an improved testing environment with more realistic tissue structures compared to traditional CT phantoms, allowing for structure-based image quality evaluation beyond noise and contrast-based assessments.

Observer studies of image quality of denoising reduced-count cardiac single photon emission computed tomography myocardial perfusion imaging by three-dimensional Gaussian post-reconstruction filtering and deep learning

BACKGROUND

The aim of this research was to asses perfusion-defect detection-accuracy by human observers as a function of reduced-counts for 3D Gaussian post-reconstruction filtering vs deep learning (DL) denoising to determine if there was improved performance with DL.

METHODS

SPECT projection data of 156 normally interpreted patients were used for these studies. Half were altered to include hybrid perfusion defects with defect presence and location known. Ordered-subset expectation-maximization (OSEM) reconstruction was employed with the optional correction of attenuation (AC) and scatter (SC) in addition to distance-dependent resolution (RC). Count levels varied from full-counts (100%) to 6.25% of full-counts. The denoising strategies were previously optimized for defect detection using total perfusion deficit (TPD). Four medical physicist (PhD) and six physician (MD) observers rated the slices using a graphical user interface. Observer ratings were analyzed using the LABMRMC multi-reader, multi-case receiver-operating-characteristic (ROC) software to calculate and compare statistically the area-under-the-ROC-curves (AUCs).

RESULTS

For the same count-level no statistically significant increase in AUCs for DL over Gaussian denoising was determined when counts were reduced to either the 25% or 12.5% of full-counts. The average AUC for full-count OSEM with solely RC and Gaussian filtering was lower than for the strategies with AC and SC, except for a reduction to 6.25% of full-counts, thus verifying the utility of employing AC and SC with RC.

CONCLUSION

We did not find any indication that at the dose levels investigated and with the DL network employed, that DL denoising was superior in AUC to optimized 3D post-reconstruction Gaussian filtering.

Unsupervised Medical Image Translation With Adversarial Diffusion Models

Imputation of missing images via source-to-target modality translation can improve diversity in medical imaging protocols. A pervasive approach for synthesizing target images involves one-shot mapping through generative adversarial networks (GAN). Yet, GAN models that implicitly characterize the image distribution can suffer from limited sample fidelity. Here, we propose a novel method based on adversarial diffusion modeling, SynDiff, for improved performance in medical image translation. To capture a direct correlate of the image distribution, SynDiff leverages a conditional diffusion process that progressively maps noise and source images onto the target image. For fast and accurate image sampling during inference, large diffusion steps are taken with adversarial projections in the reverse diffusion direction. To enable training on unpaired datasets, a cycle-consistent architecture is devised with coupled diffusive and non-diffusive modules that bilaterally translate between two modalities. Extensive assessments are reported on the utility of SynDiff against competing GAN and diffusion models in multi-contrast MRI and MRI-CT translation. Our demonstrations indicate that SynDiff offers quantitatively and qualitatively superior performance against competing baselines.

A Deep Convolutional Gated Recurrent Unit for CT Image Reconstruction

Computed tomography (CT) is one of the most important medical imaging technologies in use today. Most commercial CT products use a technique known as the filtered backprojection (FBP) that is fast and can produce decent image quality when an X-ray dose is high. However, the FBP is not good enough on low-dose X-ray CT imaging because the CT image reconstruction problem becomes more stochastic. A more effective reconstruction technique proposed recently and implemented in a limited number of CT commercial products is an iterative reconstruction (IR). The IR technique is based on a Bayesian formulation of the CT image reconstruction problem with an explicit model of the CT scanning, including its stochastic nature, and a prior model that incorporates our knowledge about what a good CT image should look like. However, constructing such prior knowledge is more complicated than it seems. In this article, we propose a novel neural network for CT image reconstruction. The network is based on the IR formulation and constructed with a recurrent neural network (RNN). Specifically, we transform the gated recurrent unit (GRU) into a neural network performing CT image reconstruction. We call it "GRU reconstruction." This neural network conducts concurrent dual-domain learning. Many deep learning (DL)-based methods in medical imaging are single-domain learning, but dual-domain learning performs better because it learns from both the sinogram and the image domain. In addition, we propose backpropagation through stage (BPTS) as a new RNN backpropagation algorithm. It is similar to the backpropagation through time (BPTT) of an RNN; however, it is tailored for iterative optimization. Results from extensive experiments indicate that our proposed method outperforms conventional model-based methods, single-domain DL methods, and state-of-the-art DL techniques in terms of the root mean squared error (RMSE), the peak signal-to-noise ratio (PSNR), and the structure similarity (SSIM) and in terms of visual appearance.

Chest tomosynthesis deblurring using CNN with deconvolution layer for vertebrae segmentation

BACKGROUND

Limited scan angles cause severe distortions and artifacts in reconstructed tomosynthesis images when the Feldkamp-Davis-Kress (FDK) algorithm is used for the purpose, which degrades clinical diagnostic performance. These blurring artifacts are fatal in chest tomosynthesis images because precise vertebrae segmentation is crucial for various diagnostic analyses, such as early diagnosis, surgical planning, and injury detection. Moreover, because most spinal pathologies are related to vertebral conditions, the development of methods for accurate and objective vertebrae segmentation in medical images is an important and challenging research area.

PURPOSE

The existing point-spread-function-(PSF)-based deblurring methods use the same PSF in all sub-volumes without considering the spatially varying property of tomosynthesis images. This increases the PSF estimation error, thus further degrading the deblurring performance. However, the proposed method estimates the PSF more accurately by using sub-CNNs that contain a deconvolution layer for each sub-system, which improves the deblurring performance.

METHODS

To minimize the effect of the spatially varying property, the proposed deblurring network architecture comprises four modules: (1) block division module, (2) partial PSF module, (3) deblurring block module, and (4) assembling block module. We compared the proposed DL-based method with the FDK algorithm, total-variation iterative reconstruction with GP-BB (TV-IR), 3D U-Net, FBPConvNet, and two-phase deblurring method. To investigate the deblurring performance of the proposed method, we evaluated its vertebrae segmentation performance by comparing the pixel accuracy (PA), intersection-over-union (IoU), and F-score values of reference images to those of the deblurred images. Also, pixel-based evaluations of the reference and deblurred images were performed by comparing their root mean squared error (RMSE) and visual information fidelity (VIF) values. In addition, 2D analysis of the deblurred images were performed by artifact spread function (ASF) and full width half maximum (FWHM) of the ASF curve.

RESULTS

The proposed method was able to recover the original structure significantly, thereby further improving the image quality. The proposed method yielded the best deblurring performance in terms of vertebrae segmentation and similarity. The IoU, F-score, and VIF values of the chest tomosynthesis images reconstructed using the proposed SV method were 53.5%, 28.7%, and 63.2% higher, respectively, than those of the images reconstructed using the FDK method, and the RMSE value was 80.3% lower. These quantitative results indicate that the proposed method can effectively restore both the vertebrae and the surrounding soft tissue.

CONCLUSIONS

We proposed a chest tomosynthesis deblurring technique for vertebrae segmentation by considering the spatially varying property of tomosynthesis systems. The results of quantitative evaluations indicated that the vertebrae segmentation performance of the proposed method was better than those of the existing deblurring methods.

Image quality improvement in bowtie-filter-equipped cone-beam CT using a dual-domain neural network

BACKGROUND

The bowtie-filter in cone-beam CT (CBCT) causes spatially nonuniform x-ray beam often leading to eclipse artifacts in the reconstructed image. The artifacts are further confounded by the patient scatter, which is therefore patient-dependent as well as system-specific.

PURPOSE

In this study, we propose a dual-domain network for reducing the bowtie-filter-induced artifacts in CBCT images.

METHODS

In the projection domain, the network compensates for the filter-induced beam-hardening that are highly related to the eclipse artifacts. The output of the projection-domain network was used for image reconstruction and the reconstructed images were fed into the image-domain network. In the image domain, the network further reduces the remaining cupping artifacts that are associated with the scatter. A single image-domain-only network was also implemented for comparison.

RESULTS

The proposed approach successfully enhanced soft-tissue contrast with much-reduced image artifacts. In the numerical study, the proposed method decreased perceptual loss and root-mean-square-error (RMSE) of the images by 84.5% and 84.9%, respectively, and increased the structure similarity index measure (SSIM) by 0.26 compared to the original input images on average. In the experimental study, the proposed method decreased perceptual loss and RMSE of the images by 87.2% and 92.1%, respectively, and increased SSIM by 0.58 compared to the original input images on average.

CONCLUSIONS

We have proposed a deep-learning-based dual-domain framework to reduce the bowtie-filter artifacts and to increase the soft-tissue contrast in CBCT images. The performance of the proposed method has been successfully demonstrated in both numerical and experimental studies.

TT U-Net: Temporal Transformer U-Net for Motion Artifact Reduction Using PAD (Pseudo All-Phase Clinical-Dataset) in Cardiac CT

Involuntary motion of the heart remains a challenge for cardiac computed tomography (CT) imaging. Although the electrocardiogram (ECG) gating strategy is widely adopted to perform CT scans at the quasi-quiescent cardiac phase, motion-induced artifacts are still unavoidable for patients with high heart rates or irregular rhythms. Dynamic cardiac CT, which provides functional information of the heart, suffers even more severe motion artifacts. In this paper, we develop a deep learning based framework for motion artifact reduction in dynamic cardiac CT. First, we build a PAD (Pseudo All-phase clinical-Dataset) based on a whole-heart motion model and single-phase cardiac CT images. This dataset provides dynamic CT images with realistic-looking motion artifacts that help to develop data-driven approaches. Second, we formulate the problem of motion artifact reduction as a video deblurring task according to its dynamic nature. A novel TT U-Net (Temporal Transformer U-Net) is proposed to excavate the spatiotemporal features for better motion artifact reduction. The self-attention mechanism along the temporal dimension effectively encodes motion information and thus aids image recovery. Experiments show that the TT U-Net trained on the proposed PAD performs well on clinical CT scans, which substantiates the effectiveness and fine generalization ability of our method. The source code, trained models, and dynamic demo will be available at https://github.com/ivy9092111111/TT-U-Net.

Synthetic Megavoltage Cone Beam Computed Tomography Image Generation for Improved Contouring Accuracy of Cardiac Pacemakers

In this study, we aimed to enhance the contouring accuracy of cardiac pacemakers by improving their visualization using deep learning models to predict MV CBCT images based on kV CT or CBCT images. Ten pacemakers and four thorax phantoms were included, creating a total of 35 combinations. Each combination was imaged on a Varian Halcyon (kV/MV CBCT images) and Siemens SOMATOM CT scanner (kV CT images). Two generative adversarial network (GAN)-based models, cycleGAN and conditional GAN (cGAN), were trained to generate synthetic MV (sMV) CBCT images from kV CT/CBCT images using twenty-eight datasets (80%). The pacemakers in the sMV CBCT images and original MV CBCT images were manually delineated and reviewed by three users. The Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and mean surface distance (MSD) were used to compare contour accuracy. Visual inspection showed the improved visualization of pacemakers on sMV CBCT images compared to original kV CT/CBCT images. Moreover, cGAN demonstrated superior performance in enhancing pacemaker visualization compared to cycleGAN. The mean DSC, HD95, and MSD for contours on sMV CBCT images generated from kV CT/CBCT images were 0.91 ± 0.02/0.92 ± 0.01, 1.38 ± 0.31 mm/1.18 ± 0.20 mm, and 0.42 ± 0.07 mm/0.36 ± 0.06 mm using the cGAN model. Deep learning-based methods, specifically cycleGAN and cGAN, can effectively enhance the visualization of pacemakers in thorax kV CT/CBCT images, therefore improving the contouring precision of these devices.

A Novel Artificial Intelligence Based Denoising Method for Ultra-Low Dose CT Used for Lung Cancer Screening

RATIONALE AND OBJECTIVES

To assess ultra-low-dose (ULD) computed tomography as well as a novel artificial intelligence-based reconstruction denoising method for ULD (dULD) in screening for lung cancer.

MATERIALS AND METHODS

This prospective study included 123 patients, 84 (70.6%) men, mean age 62.6 ± 5.35 (55-75), who had a low dose and an ULD scan. A fully convolutional-network, trained using a unique perceptual loss was used for denoising. The network used for the extraction of the perceptual features was trained in an unsupervised manner on the data itself by denoising stacked auto-encoders. The perceptual features were a combination of feature maps taken from different layers of the network, instead of using a single layer for training. Two readers independently reviewed all sets of images.

RESULTS

ULD decreased average radiation-dose by 76% (48%-85%). When comparing negative and actionable Lung-RADS categories, there was no difference between dULD and LD (p = 0.22 RE, p > 0.999 RR) nor between ULD and LD scans (p = 0.75 RE, p > 0.999 RR). ULD negative likelihood ratio (LR) for the readers was 0.033-0.097. dULD performed better with a negative LR of 0.021-0.051. Coronary artery calcifications (CAC) were documented on the dULD scan in 88(74%) and 81(68%) patients, and on the ULD in 74(62.2%) and 77(64.7%) patients. The dULD demonstrated high sensitivity, 93.9%-97.6%, with an accuracy of 91.7%. An almost perfect agreement between readers was noted for CAC scores: for LD (ICC = 0.924), dULD (ICC = 0.903), and for ULD (ICC = 0.817) scans.

CONCLUSION

A novel AI-based denoising method allows a substantial decrease in radiation dose, without misinterpretation of actionable pulmonary nodules or life-threatening findings such as aortic aneurysms.

ADDNS: An asymmetric dual deep network with sharing mechanism for medical image fusion of CT and MR-T2

Medical images with different modalities have different semantic characteristics. Medical image fusion aiming to promotion of the visual quality and practical value has become important in medical diagnostics. However, the previous methods do not fully represent semantic and visual features, and the model generalization ability needs to be improved. Furthermore, the brightness-stacking phenomenon is easy to occur during the fusion process. In this paper, we propose an asymmetric dual deep network with sharing mechanism (ADDNS) for medical image fusion. In our asymmetric model-level dual framework, primal Unet part learns to fuse medical images of different modality into a fusion image, while dual Unet part learns to invert the fusion task for multi-modal image reconstruction. This asymmetry of network settings not only enables the ADDNS to fully extract semantic and visual features, but also reduces the model complexity and accelerates the convergence. Furthermore, the sharing mechanism designed according to task relevance also reduces the model complexity and improves the generalization ability of our model. In the end, we use the intermediate supervision method to minimize the difference between fusion image and source images so as to prevent the brightness-stacking problem. Experimental results show that our algorithm achieves better results on both quantitative and qualitative experiments than several state-of-the-art methods.

Unsharp Structure Guided Filtering for Self-Supervised Low-Dose CT Imaging

Low-dose computed tomography (LDCT) imaging faces great challenges. Although supervised learning has revealed great potential, it requires sufficient and high-quality references for network training. Therefore, existing deep learning methods have been sparingly applied in clinical practice. To this end, this paper presents a novel Unsharp Structure Guided Filtering (USGF) method, which can reconstruct high-quality CT images directly from low-dose projections without clean references. Specifically, we first employ low-pass filters to estimate the structure priors from the input LDCT images. Then, inspired by classical structure transfer techniques, deep convolutional networks are adopted to implement our imaging method which combines guided filtering and structure transfer. Finally, the structure priors serve as the guidance images to alleviate over-smoothing, as they can transfer specific structural characteristics to the generated images. Furthermore, we incorporate traditional FBP algorithms into self-supervised training to enable the transformation of projection domain data to the image domain. Extensive comparisons and analyses on three datasets demonstrate that the proposed USGF has achieved superior performance in terms of noise suppression and edge preservation, and could have a significant impact on LDCT imaging in the future.

SEA-Net: Structure-Enhanced Attention Network for Limited-Angle CBCT Reconstruction of Clinical Projection Data

This work aims to improve limited-angle (LA) cone beam computed tomography (CBCT) by developing deep learning (DL) methods for real clinical CBCT projection data, which is the first feasibility study of clinical-projection-data-based LA-CBCT, to the best of our knowledge. In radiation therapy (RT), CBCT is routinely used as the on-board imaging modality for patient setup. Compared to diagnostic CT, CBCT has a long acquisition time, e.g., 60 seconds for a full 360° rotation, which is subject to the motion artifact. Therefore, the LA-CBCT, if achievable, is of the great interest for the purpose of RT, for its proportionally reduced scanning time in addition to the radiation dose. However, LA-CBCT suffers from severe wedge artifacts and image distortions. Targeting at real clinical projection data, we have explored various DL methods such as image/data/hybrid-domain methods and finally developed a so-called Structure-Enhanced Attention Network (SEA-Net) method that has the best image quality from clinical projection data among the DL methods we have implemented. Specifically, the proposed SEA-Net employs a specialized structure enhancement sub-network to promote texture preservation. Based on the observation that the distribution of wedge artifacts in reconstruction images is non-uniform, the spatial attention module is utilized to emphasize the relevant regions while ignores the irrelevant ones, which leads to more accurate texture restoration.

CARDIAN: a novel computational approach for real-time end-diastolic frame detection in intravascular ultrasound using bidirectional attention networks

Introduction

Changes in coronary artery luminal dimensions during the cardiac cycle can impact the accurate quantification of volumetric analyses in intravascular ultrasound (IVUS) image studies. Accurate ED-frame detection is pivotal for guiding interventional decisions, optimizing therapeutic interventions, and ensuring standardized volumetric analysis in research studies. Images acquired at different phases of the cardiac cycle may also lead to inaccurate quantification of atheroma volume due to the longitudinal motion of the catheter in relation to the vessel. As IVUS images are acquired throughout the cardiac cycle, end-diastolic frames are typically identified retrospectively by human analysts to minimize motion artefacts and enable more accurate and reproducible volumetric analysis.

Methods

In this paper, a novel neural network-based approach for accurate end-diastolic frame detection in IVUS sequences is proposed, trained using electrocardiogram (ECG) signals acquired synchronously during IVUS acquisition. The framework integrates dedicated motion encoders and a bidirectional attention recurrent network (BARNet) with a temporal difference encoder to extract frame-by-frame motion features corresponding to the phases of the cardiac cycle. In addition, a spatiotemporal rotation encoder is included to capture the IVUS catheter's rotational movement with respect to the coronary artery.

Results

With a prediction tolerance range of 66.7 ms, the proposed approach was able to find 71.9%, 67.8%, and 69.9% of end-diastolic frames in the left anterior descending, left circumflex and right coronary arteries, respectively, when tested against ECG estimations. When the result was compared with two expert analysts' estimation, the approach achieved a superior performance.

Discussion

These findings indicate that the developed methodology is accurate and fully reproducible and therefore it should be preferred over experts for end-diastolic frame detection in IVUS sequences.

Self-supervised denoising of projection data for low-dose cone-beam CT

BACKGROUND

Convolutional neural networks (CNNs) have shown promising results in image denoising tasks. While most existing CNN-based methods depend on supervised learning by directly mapping noisy inputs to clean targets, high-quality references are often unavailable for interventional radiology such as cone-beam computed tomography (CBCT).

PURPOSE

In this paper, we propose a novel self-supervised learning method that reduces noise in projections acquired by ordinary CBCT scans.

METHODS

With a network that partially blinds input, we are able to train the denoising model by mapping the partially blinded projections to the original projections. Additionally, we incorporate noise-to-noise learning into the self-supervised learning by mapping the adjacent projections to the original projections. With standard image reconstruction methods such as FDK-type algorithms, we can reconstruct high-quality CBCT images from the projections denoised by our projection-domain denoising method.

RESULTS

In the head phantom study, we measure peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) values of the proposed method along with the other denoising methods and uncorrected low-dose CBCT data for a quantitative comparison both in projection and image domains. The PSNR and SSIM values of our self-supervised denoising approach are 27.08 and 0.839, whereas those of uncorrected CBCT images are 15.68 and 0.103, respectively. In the retrospective study, we assess the quality of interventional patient CBCT images to evaluate the projection-domain and image-domain denoising methods. Both qualitative and quantitative results indicate that our approach can effectively produce high-quality CBCT images with low-dose projections in the absence of duplicate clean or noisy references.

CONCLUSIONS

Our self-supervised learning strategy is capable of restoring anatomical information while efficiently removing noise in CBCT projection data.

CTFlow: Mitigating Effects of Computed Tomography Acquisition and Reconstruction with Normalizing Flows

Mitigating the effects of image appearance due to variations in computed tomography (CT) acquisition and reconstruction parameters is a challenging inverse problem. We present CTFlow, a normalizing flows-based method for harmonizing CT scans acquired and reconstructed using different doses and kernels to a target scan. Unlike existing state-of-the-art image harmonization approaches that only generate a single output, flow-based methods learn the explicit conditional density and output the entire spectrum of plausible reconstruction, reflecting the underlying uncertainty of the problem. We demonstrate how normalizing flows reduces variability in image quality and the performance of a machine learning algorithm for lung nodule detection. We evaluate the performance of CTFlow by 1) comparing it with other techniques on a denoising task using the AAPM-Mayo Clinical Low-Dose CT Grand Challenge dataset, and 2) demonstrating consistency in nodule detection performance across 186 real-world low-dose CT chest scans acquired at our institution. CTFlow performs better in the denoising task for both peak signal-to-noise ratio and perceptual quality metrics. Moreover, CTFlow produces more consistent predictions across all dose and kernel conditions than generative adversarial network (GAN)-based image harmonization on a lung nodule detection task. The code is available at https://github.com/hsu-lab/ctflow.

Pathological changes or technical artefacts? The problem of the heterogenous databases in COVID-19 CXR image analysis

BACKGROUND

When the COVID-19 pandemic commenced in 2020, scientists assisted medical specialists with diagnostic algorithm development. One scientific research area related to COVID-19 diagnosis was medical imaging and its potential to support molecular tests. Unfortunately, several systems reported high accuracy in development but did not fare well in clinical application. The reason was poor generalization, a long-standing issue in AI development. Researchers found many causes of this issue and decided to refer to them as confounders, meaning a set of artefacts and methodological errors associated with the method. We aim to contribute to this steed by highlighting an undiscussed confounder related to image resolution.

METHODS

20 216 chest X-ray images (CXR) from worldwide centres were analyzed. The CXRs were bijectively projected into the 2D domain by performing Uniform Manifold Approximation and Projection (UMAP) embedding on the radiomic features (rUMAP) or CNN-based neural features (nUMAP) from the pre-last layer of the pre-trained classification neural network. Additional 44 339 thorax CXRs were used for validation. The comprehensive analysis of the multimodality of the density distribution in rUMAP/nUMAP domains and its relation to the original image properties was used to identify the main confounders.

RESULTS

nUMAP revealed a hidden bias of neural networks towards the image resolution, which the regular up-sampling procedure cannot compensate for. The issue appears regardless of the network architecture and is not observed in a high-resolution dataset. The impact of the resolution heterogeneity can be partially diminished by applying advanced deep-learning-based super-resolution networks.

CONCLUSIONS

rUMAP and nUMAP are great tools for image homogeneity analysis and bias discovery, as demonstrated by applying them to COVID-19 image data. Nonetheless, nUMAP could be applied to any type of data for which a deep neural network could be constructed. Advanced image super-resolution solutions are needed to reduce the impact of the resolution diversity on the classification network decision.

Deep learning denoising of digital breast tomosynthesis: Observer performance study of the effect on detection of microcalcifications in breast phantom images

BACKGROUND

The noise in digital breast tomosynthesis (DBT) includes x-ray quantum noise and detector readout noise. The total radiation dose of a DBT scan is kept at about the level of a digital mammogram but the detector noise is increased due to acquisition of multiple projections. The high noise can degrade the detectability of subtle lesions, specifically microcalcifications (MCs).

PURPOSE

We previously developed a deep-learning-based denoiser to improve the image quality of DBT. In the current study, we conducted an observer performance study with breast radiologists to investigate the feasibility of using deep-learning-based denoising to improve the detection of MCs in DBT.

METHODS

We have a modular breast phantom set containing seven 1-cm-thick heterogeneous 50% adipose/50% fibroglandular slabs custom-made by CIRS, Inc. (Norfolk, VA). We made six 5-cm-thick breast phantoms embedded with 144 simulated MC clusters of four nominal speck sizes (0.125-0.150, 0.150-0.180, 0.180-0.212, 0.212-0.250 mm) at random locations. The phantoms were imaged with a GE Pristina DBT system using the automatic standard (STD) mode. The phantoms were also imaged with the STD+ mode that increased the average glandular dose by 54% to be used as a reference condition for comparison of radiologists' reading. Our previously trained and validated denoiser was deployed to the STD images to obtain a denoised DBT set (dnSTD). Seven breast radiologists participated as readers to detect the MCs in the DBT volumes of the six phantoms under the three conditions (STD, STD+, dnSTD), totaling 18 DBT volumes. Each radiologist read all the 18 DBT volumes sequentially, which were arranged in a different order for each reader in a counter-balanced manner to minimize any potential reading order effects. They marked the location of each detected MC cluster and provided a conspicuity rating and their confidence level for the perceived cluster. The visual grading characteristics (VGC) analysis was used to compare the conspicuity ratings and the confidence levels of the radiologists for the detection of MCs.

RESULTS

The average sensitivities over all MC speck sizes were 65.3%, 73.2%, and 72.3%, respectively, for the radiologists reading the STD, dnSTD, and STD+ volumes. The sensitivity for dnSTD was significantly higher than that for STD (p < 0.005, two-tailed Wilcoxon signed rank test) and comparable to that for STD+. The average false positive rates were 3.9 ± 4.6, 2.8 ± 3.7, and 2.7 ± 3.9 marks per DBT volume, respectively, for reading the STD, dnSTD, and STD+ images but the difference between dnSTD and STD or STD+ did not reach statistical significance. The overall conspicuity ratings and confidence levels by VGC analysis for dnSTD were significantly higher than those for both STD and STD+ (p ≤ 0.001). The critical alpha value for significance was adjusted to be 0.025 with Bonferroni correction.

CONCLUSIONS

This observer study using breast phantom images showed that deep-learning-based denoising has the potential to improve the detection of MCs in noisy DBT images and increase radiologists' confidence in differentiating noise from MCs without increasing radiation dose. Further studies are needed to evaluate the generalizability of these results to the wide range of DBTs from human subjects and patient populations in clinical settings.

A cascade-based dual-domain data correction network for sparse view CT image reconstruction

Computed tomography (CT) provides non-invasive anatomical structures of the human body and is also widely used for clinical diagnosis, but excessive ionizing radiation in X-rays can cause harm to the human body. Therefore, the researchers obtained sparse sinograms reconstructed sparse view CT images (SVCT) by reducing the amount of X-ray projection, thereby reducing the radiological effects caused by radiation. This paper proposes a cascade-based dual-domain data correction network (CDDCN), which can effectively combine the complementary information contained in the sinogram domain and the image domain to reconstruct high-quality CT images from sparse view sinograms. Specifically, several encoder-decoder subnets are cascaded in the sinogram domain to reconstruct artifact-free and noise-free CT images. In the encoder-decoder subnets, spatial-channel domain learning is designed to achieve efficient feature fusion through a group merging structure, providing continuous and elaborate pixel-level features and improving feature extraction efficiency. At the same time, to ensure that the original sinogram data collected can be retained, a sinogram data consistency layer is proposed to ensure the fidelity of the sinogram data. To further maintain the consistency between the reconstructed image and the reference image, a multi-level composite loss function is designed for regularization to compensate for excessive smoothing and distortion of the image caused by pixel loss and preserve image details and texture. Quantitative and qualitative analysis shows that CDDCN achieves competitive results in artifact removal, edge preservation, detail restoration, and visual improvement for sparsely sampled data under different views.

HCformer: Hybrid CNN-Transformer for LDCT Image Denoising

Low-dose computed tomography (LDCT) is an effective way to reduce radiation exposure for patients. However, it will increase the noise of reconstructed CT images and affect the precision of clinical diagnosis. The majority of the current deep learning-based denoising methods are built on convolutional neural networks (CNNs), which concentrate on local information and have little capacity for multiple structures modeling. Transformer structures are capable of computing each pixel's response on a global scale, but their extensive computation requirements prevent them from being widely used in medical image processing. To reduce the impact of LDCT scans on patients, this paper aims to develop an image post-processing method by combining CNN and Transformer structures. This method can obtain a high-quality images from LDCT. A hybrid CNN-Transformer (HCformer) codec network model is proposed for LDCT image denoising. A neighborhood feature enhancement (NEF) module is designed to introduce the local information into the Transformer's operation, and the representation of adjacent pixel information in the LDCT image denoising task is increased. The shifting window method is utilized to lower the computational complexity of the network model and overcome the problems that come with computing the MSA (Multi-head self-attention) process in a fixed window. Meanwhile, W/SW-MSA (Windows/Shifted window Multi-head self-attention) is alternately used in two layers of the Transformer to gain the information interaction between various Transformer layers. This approach can successfully decrease the Transformer's overall computational cost. The AAPM 2016 LDCT grand challenge dataset is employed for ablation and comparison experiments to demonstrate the viability of the proposed LDCT denoising method. Per the experimental findings, HCformer can increase the image quality metrics SSIM, HuRMSE and FSIM from 0.8017, 34.1898, and 0.6885 to 0.8507, 17.7213, and 0.7247, respectively. Additionally, the proposed HCformer algorithm will preserves image details while it reduces noise. In this paper, an HCformer structure is proposed based on deep learning and evaluated by using the AAPM LDCT dataset. Both the qualitative and quantitative comparison results confirm that the proposed HCformer outperforms other methods. The contribution of each component of the HCformer is also confirmed by the ablation experiments. HCformer can combine the advantages of CNN and Transformer, and it has great potential for LDCT image denoising and other tasks.

Measurement of proximal contact of single crowns to assess interproximal relief: A pilot study

Background

It is common for dental technicians to adjust the proximal surface of adjacent teeth on casts when fabricating single crowns. However, whether the accuracy of the proximal contact is affected if this step is eliminated is unclear.

Objective

To evaluate the accuracy of the proximal contact of single crowns for mandibular first molars fabricated from four different restorative materials, without adjustment of the proximal surface of the adjacent teeth by the laboratory/dental technician.

Methods

This study was in vitro; all the clinical procedures were conducted on a dentoform. The mandibular first molar tooth on the dentoform was prepared using diamond burs and a high speed handpiece. Twenty single crowns were fabricated, five for each group (monolithic zirconia, lithium disilicate, metal ceramic, and cast gold). No proximal surface adjacent to the definitive crowns was adjusted for tight contact in the dental laboratory. Both the qualitative analyses, using dental floss and shimstock, and the quantitative analyses, using a stereo microscope, were performed to evaluate the accuracy of the proximal contact of the restoration with the adjacent teeth. In the quantitative analysis, one-way analysis of variance was used to compare mean values at a significance level of 0.05.

Results

In quantitative analysis, the differences between the proximal contact tightness of the four groups was not statistically significant (P = 0.802 for mesial contacts, P = 0.354 for distal contacts). In qualitative analysis, in most crowns, dental floss passed through the contact with tight resistance and only one film of shimstock could be inserted between the adjacent teeth and the restoration. However, one specimen from the cast gold crown had open contact.

Conclusions

Even without proximal surface adjustment of the adjacent teeth during the crown fabrication process, adequate proximal contact tightness between the restoration and adjacent teeth could be achieved.

Self-supervised deep learning for joint 3D low-dose PET/CT image denoising

Deep learning (DL)-based denoising of low-dose positron emission tomography (LDPET) and low-dose computed tomography (LDCT) has been widely explored. However, previous methods have focused only on single modality denoising, neglecting the possibility of simultaneously denoising LDPET and LDCT using only one neural network, i.e., joint LDPET/LDCT denoising. Moreover, DL-based denoising methods generally require plenty of well-aligned LD-normal-dose (LD-ND) sample pairs, which can be difficult to obtain. To this end, we propose a self-supervised two-stage training framework named MAsk-then-Cycle (MAC), to achieve self-supervised joint LDPET/LDCT denoising. The first stage of MAC is masked autoencoder (MAE)-based pre-training and the second stage is self-supervised denoising training. Specifically, we propose a self-supervised denoising strategy named cycle self-recombination (CSR), which enables denoising without well-aligned sample pairs. Unlike other methods that treat noise as a homogeneous whole, CSR disentangles noise into signal-dependent and independent noises. This is more in line with the actual imaging process and allows for flexible recombination of noises and signals to generate new samples. These new samples contain implicit constraints that can improve the network's denoising ability. Based on these constraints, we design multiple loss functions to enable self-supervised training. Then we design a CSR-based denoising network to achieve joint 3D LDPET/LDCT denoising. Existing self-supervised methods generally lack pixel-level constraints on networks, which can easily lead to additional artifacts. Before denoising training, we perform MAE-based pre-training to indirectly impose pixel-level constraints on networks. Experiments on an LDPET/LDCT dataset demonstrate its superiority over existing methods. Our method is the first self-supervised joint LDPET/LDCT denoising method. It does not require any prior assumptions and is therefore more robust.