Low-dose CT Images Research Articles

MLAR-UNet: LDCT image denoising based on U-Net with multiple lightweight attention-based modules and residual reinforcement.

Computed tomography (CT) is a crucial medical imaging technique which uses X-ray radiation to identify cancer tissues. Since radiation poses a significant health risk, low dose acquisition procedures need to be adopted. However, low-dose CT (LDCT) can cause higher noise and artifacts which massively degrade the diagnosis. To denoise LDCT images more effectively, this paper proposes a deep learning method based on U-Net with multiple lightweight attention-based modules and residual reinforcement (MLAR-UNet), We integrate a U-Net architecture with several advanced modules, including Convolutional Block Attention Module (CBAM), Cross Residual Module (CR), Attention Cross Reinforcement Module (ACRM), and Convolution and Transformer Cross Attention Module (CTCAM). Among these modules, CBAM applies channel and spatial attention mechanisms to enhance local feature representation. However, serious detail loss caused by incorrect embedding of CBAM for LDCT denoising is verified in this study. To relieve this, we introduce CR to reduce information loss in deeper layers, preserving features more effectively. To address the excessive local attention of CBAM, we design ACRM, which incorporates Transformer to adjust the attention weights. Furthermore, we design CTCAM, which leverages a complex combination of Transformer and convolution to capture multi-scale information and compute more accurate attention weights. Experiments verify the embedding rationality and validity of each module and show that the proposed MLAR-UNet denoises LDCT images more effectively and preserves more details than many state-of-the-art (SOTA) methods on clinical chest and abdominal CT datasets. The proposed MLAR-UNet not only demonstrates superior LDCT image denoising capability but also highlights the strong detail comprehension and negligible overheads of our designed ACRM and CTCAM. These findings provide a novel approach to integrating Transformer more efficiently in image processing.

Concordance-based Predictive Uncertainty (CPU)-Index: Proof-of-concept with application towards improved specificity of lung cancers on low dose screening CT.

In this paper, we introduce a novel concordance-based predictive uncertainty (CPU)-Index, which integrates insights from subgroup analysis and personalized AI time-to-event models. Through its application in refining lung cancer screening (LCS) predictions generated by an individualized AI time-to-event model trained with fused data of low dose CT (LDCT) radiomics with patient demographics, we demonstrate its effectiveness, resulting in improved risk assessment compared to the Lung CT Screening Reporting & Data System (Lung-RADS). Subgroup-based Lung-RADS faces challenges in representing individual variations and relies on a limited set of predefined characteristics, resulting in variable predictions. Conversely, personalized AI time-to-event models are hindered by transparency issues and biases from censored data. By measuring the prediction consistency between subgroup analysis and AI time-to-event models, the CPU-Index framework offers a nuanced evaluation of the bias-variance trade-off and improves the transparency and reliability of predictions. Consistency was estimated by the concordance index of subgroup analysis-based similarity rank and model prediction similarity rank. Subgroup analysis-based similarity loss was defined as the sum-of-the-difference between Lung-RADS and feature-level 0-1 loss. Model prediction similarity loss was defined as squared loss. To test our approach, we identified 3,326 patients who underwent LDCT for LCS from 1/1/2015 to 6/30/2020 with confirmation of lung cancer on pathology within one year. For each LDCT image, the lesion associated with a Lung-RADS score was detected using a pretrained deep learning model from Medical Open Network for AI (MONAI), from which radiomic features were extracted. Radiomics were optimally fused with patient demographics via a positional encoding scheme and used to train a neural multi-task logistic regression time-to-event model that predicts malignancy. Performance was maximized when radiomics features were fused with positionally encoded demographic features. In this configuration, our algorithm raised the AUC from 0.81 ± 0.04 to 0.89 ± 0.02. Compared to standard Lung-RADS, our approach reduced the False-Positive-Rate from 0.41 ± 0.02 to 0.30 ± 0.12 while maintaining the same False-Negative-Rate. Our methodology enhances lung cancer risk assessment by estimating prediction uncertainty and adjusting accordingly. Furthermore, the optimal integration of radiomics and patient demographics improved overall diagnostic performance, indicating their complementary nature.

ICU-Net: A U-shaped Low-Dose CT Image Denoising Network Based on Codec Structure

ICU-Net: A U-shaped Low-Dose CT Image Denoising Network Based on Codec Structure

An early lung cancer diagnosis model for non-smokers incorporating ct imaging analysis and circulating genetically abnormal cells (CACs)

BackgroundAn increase in the prevalence of lung cancer that is not smoking-related has been noticed in recent years. Unfortunately, these patients are not included in low dose computer tomography (LDCT) screening programs and are not actually considered in early diagnosis. Therefore, improved early diagnosis methods are urgently needed for non-smokers. It is necessary to establish a prediction model for non-smoking individuals at intermediate to high risk of developing lung cancer (LC) and develop a tool to address the significant gap in evaluating pulmonary nodules in non-smokers.MethodsWe retrospectively investigated 1121 patients with pulmonary nodules, who underwent LDCT examinations between September 2019 and March 2023. Five artificial intelligence (AI) algorithms were used to build two kinds of models and identify which one was better at diagnosing non-smoking pulmonary nodules patients. In the first model, we assigned 554 non-smoking individuals to a training cohort and 150 non-smoking patients to an independent validation cohort. The second model included 971 patients for the training set and 150 non-smoking patients for an independent validation set. All LDCT images of participants were obtained for AI analysis. AI of LDCT scans, liquid biopsy, and clinical characteristics were collected for model building.ResultsAmong LC patients, 58,4% were non-smokers. Non-smoking patients had a high incidence of LC (71.4%), and women showed a significant excess risk compared with non-smoking men in terms of LC risk. Furthermore, our results indicated that the model built using random forest (RF) method, which integrates clinical characteristics (age, extra-thoracic cancer history, gender), radiological characteristics of pulmonary nodules (nodule diameter, nodule count, upper lobe location, malignant sign at the nodule edge, subsolid status), the artificial intelligence analysis of LDCT data, and liquid biopsy achieved the best diagnostic performance in the independent external non-smokers validation cohort (sensitivity 92%, specificity 97%, area under the curve [AUC] = 0.99).ConclusionsThese results could significantly improve early non-smoker LC diagnosis and treatment for non-smoker patients with malignant nodules. The established multi-omics model is a noninvasive prediction tool for non-smoking malignant pulmonary nodule diagnosis. Validation revealed that these models exhibited excellent discrimination and calibration capacities, especially the first model built using the RF method, suggesting their clinical utility in the early screening and diagnosis of non-smoking LC.

A low-dose CT reconstruction method using sub-pixel anisotropic diffusion.

We present a new low-dose CT reconstruction method using sub-pixel and anisotropic diffusion. The sub-pixel intensity values and their second-order differences were obtained using linear interpolation techniques, and the new gradient information was then embedded into an anisotropic diffusion process, which was introduced into a penalty-weighted least squares model to reduce the noise in low-dose CT projection data. The high-quality CT image was finally reconstructed using the classical filtered back-projection (FBP) algorithm from the estimated data. In the Shepp-Logan phantom experiments, the structural similarity (SSIM) index of the CT image reconstructed by the proposed algorithm, as compared with FBP, PWLS-Gibbs and PWLS-TV algorithms, was increased by 28.13%, 5.49%, and 0.91%, the feature similarity (FSIM) index was increased by 21.08%, 1.78%, and 1.36%, and the root mean square error (RMSE) was reduced by 69.59%, 18.96%, and 3.90%, respectively. In the digital XCAT phantom experiments, the SSIM index of the CT image reconstructed by the proposed algorithm, as compared with FBP, PWLS-Gibbs and PWLS-TV algorithms, was increased by 14.24%, 1.43% and 7.89%, the FSIM index was increased by 9.61%, 1.78% and 5.66%, and the RMSE was reduced by 26.88%, 9.41% and 18.39%, respectively. In clinical experiments, the SSIM index of the image reconstructed using the proposed algorithm was increased by 19.24%, 15.63% and 3.68%, the FSIM index was increased by 4.30%, 2.92% and 0.43%, and the RMSE was reduced by 44.60%, 36.84% and 15.22% in comparison with FBP, PWLS-Gibbs and PWLS-TV algorithms, respectively. The proposed method can effectively reduce the noises and artifacts while maintaining the structural details in low-dose CT images.

Dual-domain Wasserstein Generative Adversarial Network with Hybrid Loss for Low-dose CT Imaging

Objective.Low-dose computed tomography (LDCT) has gained significant attention in hospitals and clinics as a popular imaging modality for reducing the risk of x-ray radiation. However, reconstructed LDCT images often suffer from undesired noise and artifacts, which can negatively impact diagnostic accuracy. This study aims to develop a novel approach to improve LDCT imaging performance.Approach.A dual-domain Wasserstein generative adversarial network (DWGAN) with hybrid loss is proposed as an effective and integrated deep neural network (DNN) for LDCT imaging. The DWGAN comprises two key components: a generator (G) network and a discriminator (D) network. TheGnetwork is a dual-domain DNN designed to predict high-quality images by integrating three essential components: the projection-domain denoising module, filtered back-projection-based reconstruction layer, and image-domain enhancement module. TheDnetwork is a shallow convolutional neural network used to differentiate between real (label) and generated images. To prevent the reconstructed images from becoming excessively smooth and to preserve both structural and textural details, a hybrid loss function with weighting coefficients is incorporated into the DWGAN.Main results.Numerical experiments demonstrate that the proposed DWGAN can effectively suppress noise and better preserve image details compared with existing methods. Moreover, its application to head CT data confirms the superior performance of the DWGAN in restoring structural and textural details.Significance.The proposed DWGAN framework exhibits excellent performance in recovering structural and textural details in LDCT images. Furthermore, the framework can be applied to other tomographic imaging techniques that suffer from image distortion problems.

A Novel Attention-Guided Enhanced U-Net with Hybrid Edge-Preserving Structural Loss for Low-Dose CT Image Denoising

A Novel Attention-Guided Enhanced U-Net with Hybrid Edge-Preserving Structural Loss for Low-Dose CT Image Denoising

Radiomics models based on thoracic and upper lumbar spine in chest LDCT to predict low bone mineral density

This study aims to develop and validate different radiomics models based on thoracic and upper lumbar spine in chest low-dose computed tomography (LDCT) to predict low bone mineral density (BMD) using quantitative computed tomography (QCT) as standard of reference. A total of 905 participants underwent chest LDCT and paired QCT BMD examination were retrospectively included from August 2018 and June 2019. The patients with low BMD (n = 388) and the normal (n = 517) were randomly divided into a training set (n = 622) and a validation set (n = 283). Radiomics features (RFs) were extracted from the single and consecutive vertebrae in chest LDCT images to construct the single vertebra RFs models, mixed RFs models and Radscore models, respectively. The performance of these models was evaluated by the area under the curve (AUC) of receiver operator characteristic curve, using QCT as standard of reference. The Radscore models, mixed RFs models, and single vertebra RFs models yielded the AUC values ranging from 0.809 to 0.906, 0.792 to 0.883, and 0.731 to 0.884 for predicting low BMD in the validation set, respectively. For predicting low BMD, the Radscore model of L1-L2 vertebrae yielded the highest AUC of 0.906, and of T1-T3 yielded the lowest AUC of 0.809 (P < 0.05), respectively. However, there was no significant difference among the AUC values of three Radscore models constructed on the vertebrae of T4-T6 (AUC = 0.855), T7-T9 (AUC = 0.845), and T10-T12 (AUC = 0.871) for predicting low BMD in the validation set (P > 0.1). The Radscore model of L1-L2 have potential to serve as an important tool for predicting and screening low BMD from normal in chest LDCT images.

Missed Opportunities for Lung Cancer Screening Among Patients With Behavioral Health Disorders With Elevated Cigarette Smoking Rates: Lung Cancer Screening and Behavioral Health.

Annual lung cancer screening using low-dose CT (LDCT) imaging effectively reduces mortality from lung cancer and is recommended for people who are at high risk of developing the disease. The utilization of lung cancer screening, however, has remained low. Due to significantly higher cigarette smoking rates, patients with behavioral health disorders (those living with mental illness and/or substance use disorders) are more likely to be diagnosed with and die of lung cancer; at the same time, they are less likely to undergo cancer screenings. There is an urgent need for targeted efforts to improve access to lung cancer screening among this population disproportionately affected by the disease. In this commentary, we propose integrating lung cancer screening facilitation into services provided by behavioral health professionals who are uniquely positioned to reach these patients and deliver interventions to increase uptake of cancer screenings. We suggest several measures that could improve lung health outcomes of patients with behavioral health disorders: (1) training behavioral health professionals in lung cancer screening eligibility assessment; (2) providing patients with educational materials; (3) integrating shared decision-making counseling for lung cancer screening into behavioral health care settings; (4) providing the practical support needed to access screening; and (5) establishing effective partnerships with community organizations. Regardless of the level of engagement, possibly ranging from brief training to the implementation of comprehensive programs, any involvement will benefit patients. This integrated approach will contribute to reducing lung cancer mortality among patients with behavioral health disorders who have long experienced systemic health inequities.

DDoCT: Morphology preserved dual-domain joint optimization for fast sparse-view low-dose CT imaging.

Computed tomography (CT) is continuously becoming a valuable diagnostic technique in clinical practice. However, the radiation dose exposure in the CT scanning process is a public health concern. Within medical diagnoses, mitigating the radiation risk to patients can be achieved by reducing the radiation dose through adjustments in tube current and/or the number of projections. Nevertheless, dose reduction introduces additional noise and artifacts, which have extremely detrimental effects on clinical diagnosis and subsequent analysis. In recent years, the feasibility of applying deep learning methods to low-dose CT (LDCT) imaging has been demonstrated, leading to significant achievements. This article proposes a dual-domain joint optimization LDCT imaging framework (termed DDoCT) which uses noisy sparse-view projection to reconstruct high-performance CT images with joint optimization in projection and image domains. The proposed method not only addresses the noise introduced by reducing tube current, but also pays special attention to issues such as streak artifacts caused by a reduction in the number of projections, enhancing the applicability of DDoCT in practical fast LDCT imaging environments. Experimental results have demonstrated that DDoCT has made significant progress in reducing noise and streak artifacts and enhancing the contrast and clarity of the images.

Serum miR-155–5p assists the diagnostic sensitivity and accuracy of low-dose spiral CT imaging in early lung cancer

BackgroundLow-dose spiral CT imaging has been employed for cancer diagnosis, but its sensitivity and specificity were unsatisfactory. MicroRNAs have been considered an approach for screening cancers, and the function of miR-155–5p in lung cancer has been previously revealed. ObjectivesThis study assessed the diagnostic value of combining low-dose spiral CT with serum miR-155–5p in lung cancer aiming to explore a novel strategy to assist the clinical cancer diagnosis. MethodsThis study enrolled 115 lung cancer patients and 115 patients with benign lung diseases as control. All patients received low-dose spiral CT imaging, and serum miR-155–5p levels were analyzed by PCR. The diagnostic potential of miR-155–5p in lung cancer was evaluated by receiver operating curve (ROC), and its significance in evaluating the risk of lung cancer was evaluated by logistic regression analysis. The consistency of serum miR-155–5p and low-dose spiral CT imaging with pathological examination was assessed by the Kappa test. ResultsReduced serum miR-155–5p indicated the risk of lung cancer in patients with benign lung diseases. Decreased serum miR-155–5p showed significant diagnostic value in early lung cancer and was significantly associated with disease severity. Low-dose spiral CT imaging showed significant diagnostic value in early lung cancer and showed middle consistency with pathological examination. Combining low-dose spiral CT imaging with serum miR-155–5p improved the diagnostic sensitivity and accuracy in early lung cancer, reduced the false positive rate, and showed better consistency with pathological examination. ConclusionSerum miR-155–5p levels could assist the early detection of lung cancer by low-dose spiral CT examination.

A Hybrid U-Net Framework for Low-Dose CT Image Denoising: Leveraging EdgeNet+ for Structural Integrity and Image Quality

A Hybrid U-Net Framework for Low-Dose CT Image Denoising: Leveraging EdgeNet+ for Structural Integrity and Image Quality

Study of Low-Dose CT Image Enhancement Network Based on Improved Cycle GAN Method

Study of Low-Dose CT Image Enhancement Network Based on Improved Cycle GAN Method

Application of Adaptive Search Window-Based Nonlocal Total Variation Filter in Low-Dose Computed Tomography Images: A Phantom Study

Computed tomography (CT) imaging using low-dose radiation effectively reduces radiation exposure; however, it introduces noise amplification in the resulting image. This study models an adaptive nonlocal total variation (NL-TV) algorithm that efficiently reduces noise in X-ray-based images and applies it to low-dose CT images. In this study, an AAPM CT performance phantom is used, and the resulting image is obtained by applying an annotation filter and a high-pitch protocol. The adaptive NL-TV filter was designed by applying the optimal window value calculated by confirming the difference between Gaussian filtering and the basic NL-TV approach. For quantitative image quality evaluation parameters, contrast-to-noise ratio (CNR), coefficient of variation (COV), and sigma value were used to confirm the noise reduction effectiveness and spatial resolution value. The CNR and COV values in low-dose CT images using the adaptive NL-TV filter, which performed an optimization process, improved by approximately 1.29 and 1.45 times, respectively, compared with conventional NL-TV. In addition, the adaptive NL-TV filter was able to acquire spatial resolution data that were similar to a CT image without applying noise reduction. In conclusion, the proposed NL-TV filter is feasible and effective in improving the quality of low-dose CT images.

Hybrid plug-and-play CT image restoration using nonconvex low-rank group sparsity and deep denoiser priors

Objective. Low-dose computed tomography (LDCT) is an imaging technique that can effectively help patients reduce radiation dose, which has attracted increasing interest from researchers in the field of medical imaging. Nevertheless, LDCT imaging is often affected by a large amount of noise, making it difficult to clearly display subtle abnormalities or lesions. Therefore, this paper proposes a multiple complementary priors CT image reconstruction method by simultaneously considering both the internal prior and external image information of CT images, thereby enhancing the reconstruction quality of CT images.Approach. Specifically, we propose a CT image reconstruction method based on weighted nonconvex low-rank regularized group sparse and deep image priors under hybrid plug-and-play framework by utilizing the weighted nonconvex low rankness and group sparsity of dictionary domain coefficients of each group of similar patches, and a convolutional neural network denoiser. To make the proposed reconstruction problem easier to tackle, we utilize the alternate direction method of multipliers for optimization.Main results. To verify the performance of the proposed method, we conduct detailed simulation experiments on the images of the abdominal, pelvic, and thoracic at projection views of 45, 65, and 85, and at noise levels of1×105and1×106, respectively. A large number of qualitative and quantitative experimental results indicate that the proposed method has achieved better results in texture preservation and noise suppression compared to several existing iterative reconstruction methods.Significance. The proposed method fully considers the internal nonlocal low rankness and sparsity, as well as the external local information of CT images, providing a more effective solution for CT image reconstruction. Consequently, this method enables doctors to diagnose and treat diseases more accurately by reconstructing high-quality CT images.

Evaluation of SR-DLR in low-dose abdominal CT: superior image quality and noise reduction.

To evaluate the effectiveness of super-resolution deep learning reconstruction (SR-DLR) in low-dose abdominal computed tomography (CT) imaging compared with hybrid iterative reconstruction (HIR) and conventional deep learning reconstruction (cDLR) algorithms. We retrospectively analyzed abdominal CT scans performed using a low-dose protocol. Three different image reconstruction algorithms-HIR, cDLR, and SR-DLR-were applied to the same raw image data. Objective evaluations included noise magnitude and contrast-to-noise ratio (CNR), as well as noise power spectrum (NPS) and edge rise slope (ERS). Subjective evaluations were performed by radiologists, who assessed image quality in terms of noise, artifacts, sharpness, and overall diagnostic utility. Raw CT image data were obtained from 35 patients (mean CTDIvol 11.0 mGy; mean DLP 344.8 mGy/cm). cDLR yielded the lowest noise levels and highest CNR (p < 0.001). However, SR-DLR outperformed cDLR in terms of noise texture and resolution, achieving the lowest NPS peak and highest ERS (p < 0.001 and p = 0.005, respectively). Subjectively, SR-DLR was rated highest across all categories, including noise, artifacts, sharpness, and overall image quality, with statistically significant differences compared to cDLR and HIR (p < 0.001). SR-DLR was the most effective reconstruction algorithm for low-dose abdominal CT imaging, offering superior image quality and noise reduction compared to cDLR and HIR. This suggests that SR-DLR can enhance the reliability and diagnostic accuracy of abdominal imaging, particularly in low-dose settings, making it a valuable tool in clinical practice.

MoCoDiff: Momentum context diffusion model for low-dose CT denoising

Low-Dose Computed Tomography (LDCT) has gradually replaced Normal-Dose Computed Tomography (NDCT) due to its lower radiation exposure. However, the reduction in radiation dose has led to increased noise and artifacts in LDCT images. To date, many methods for LDCT denoising have emerged, but they often struggle to balance denoising performance with reconstruction efficiency. This paper presents a novel Momentum Context Diffusion model for low-dose CT denoising, termed MoCoDiff. First, MoCoDiff employs a Mean-Preserving Stochastic Degradation (MPSD) operator to gradually degrade NDCT to LDCT, effectively simulating the physical process of CT degradation and greatly reducing sampling steps. Furthermore, the stochastic nature of the MPSD operator enhances the diversity of samples in the training space and calibrates the deviation between network inputs and time-step embedded features. Second, we propose a Momentum Context (MoCo) strategy. This strategy uses the most recent sampling result from each step to update the context information, thereby narrowing the noise level gap between the sampling results and the context data. This approach helps to better guide the next sampling step. Finally, to prevent issues such as over-smoothing of image edges that can arise from using the mean square error loss function, we develop a dual-domain loss function that operates in both the image and wavelet domains. This approach leverages wavelet domain information to encourage the model to preserve structural details in the images more effectively. Extensive experimental results show that our MoCoDiff model outperforms competing methods in both denoising and generalization performance, while also ensuring fast training and inference.

Role and Impact of Method Noise on CT Image Denoising

Background The main emphasis of this study is on the medical Computed Tomography (CT) imaging denoising technique, which plays a major role in interpreting patient illness information for medical diagnosis. CT imaging is indispensable for accurate disease diagnosis, but image quality is affected by noise and other artifacts. The primary objective is to improve the accuracy of denoising algorithms, which consequently increases early disease prediction and enhances the accuracy of diagnostic outcomes. Objective The major objective was to examine the performance of method noise-based Low-dose CT (LDCT) image denoising technique using a Convolutional Neural Network (CNN) in medical imaging. This method aims to suppress noise, improve image quality, and effectively minimize radiation exposure. This method enhances the accuracy of the denoising algorithm, enabling more precise disease diagnosis. Method noise, or residual noise, plays a major role in denoising CT images while preserving fine details and minimizing other artifacts generated during the denoising process. Method noise includes the omitted structural features and other minute artifacts, which are resolved through CNN-based denoising techniques. This approach elevates the overall imaging quality and clarity, resulting in better diagnostic accuracy. Methods The study includes a systematic, method noise-based approach to determine the performance of denoising algorithms in diagnosing various diseases from medical CT images that are often affected by Gaussian noise. It involves selecting a comprehensive dataset, applying a method noise approach using CNN, and evaluating the outcomes through quantitative measures, such as PSNR, SNR, and SSIM. This comparison aims to assess diagnostic interpretation, thereby improving the accuracy and efficacy of the method noise-based technique in CT medical imaging. Results The results illustrate the differential accuracy and performance of CT image denoising techniques when compared to standard filtering methods, as well as after the application of method noise-based denoising techniques. Implementing quantitative measures, such as PSNR, SNR, and SSIM, aims to improve healthcare diagnostics. Conclusion The study concludes that method noise-based CT image denoising algorithms effectively mitigate noise and artifacts while retaining the corners, contours, and precise details of CT images, subsequently improving the efficiency and accuracy of predicting diagnostic results.

Combination of edge enhancement and cold diffusion model for low dose CT image denoising.

Since the quality of low dose CT (LDCT) images is often severely affected by noise and artifacts, it is very important to maintain high quality CT images while effectively reducing the radiation dose. In recent years, the representation of diffusion models to produce high quality images and stable trainability has attracted wide attention. With the extension of the cold diffusion model to the classical diffusion model, its application has greater flexibility. Inspired by the cold diffusion model, we proposes a low dose CT image denoising method, called CECDM, based on the combination of edge enhancement and cold diffusion model. The LDCT image is taken as the end point (forward) of the diffusion process and the starting point (reverse) of the sampling process. Improved sobel operator and Convolution Block Attention Module are added to the network, and compound loss function is adopted. The experimental results show that CECDM can effectively remove noise and artifacts from LDCT images while the inference time of a single image is reduced to 0.41 s. Compared with the existing LDCT image post-processing methods, CECDM has a significant improvement in all indexes.

Low Dose CT Image Denoising: A Comparative Study of Deep Learning Models and Training Strategies

Article Low Dose CT Image Denoising: A Comparative Study of Deep Learning Models and Training Strategies Heng Zhao 1, Like Qian 1, Yaqi Zhu 1 and Dingcheng Tian 1,2,∗ 1 Research Institute for Medical and Biological Engineering, Ningbo University, Ningbo 315211, China 2 College of Medicine and Biological Information Engineering, Northeastern University, Shenyang 110016, China ∗ Correspondence: 2310520@stu.neu.edu.cn Received: 8 August 2024; Revised: 10 October 2024; Accepted: 14 October 2024; Published: 5 November 2024 Abstract: Low-dose computed tomography (LDCT) denoising is an important topic in CT image research. Compared with normal-dose CT images, LDCT can reduce the radiation dose of X-rays, decreasing the radiation burden on the human body, which is beneficial to human health. However, quantum noise caused by low-dose rays will reduce the quality of CT images, thereby decreasing the accuracy of clinical diagnosis. In recent years, deep learning-based denoising methods have shown promising advantages in this field. Researchers have proposed some optimized models for low-dose CT image denoising. These methods have enhanced the application of low-dose CT image denoising from different aspects. From the perspective of experimental research, this paper investigates and evaluates some top deep learning models proposed in the field of low-dose image denoising in recent years, with the aim of determining the best models and training strategies for this task. We conducted experiments on seven deep learning models (REDCNN, EDCNN, QAE, OCTNet, UNet, WGAN, CTformer) on the AAPM dataset and the Piglet dataset. Our research shows that UNet has the best denoising effect among the models, obtaining PSNR = 33.06 (AAPM dataset) and PSNR = 31.21 (Piglet dataset), and good generalization capacity is also observed. However, UNet has a large number of parameters, and the time it takes to process an image is about 8 ms, while EDCNN takes about 4.8 ms to process an image, and its average PSNR is ranked second after UNet. EDCNN strikes a balance between denoising performance and processing efficiency, making it ideal for low-dose CT image denoising tasks.