Deep Learning for Low-Dose CT Denoising Using Perceptual Loss and Edge Detection Layer.

Low-dose CT denoising is a challenging task that has been studied by many researchers. Some studies have used deep neural networks to improve the quality of low-dose CT images and achieved fruitful results. In this paper, we propose a deep neural network that uses dilated convolutions with different dilation rates instead of standard convolution helping to capture more contextual information in fewer layers. Also, we have employed residual learning by creating shortcut connections to transmit image information from the early layers to later ones. To further improve the performance of the network, we have introduced a non-trainable edge detection layer that extracts edges in horizontal, vertical, and diagonal directions. Finally, we demonstrate that optimizing the network by a combination of mean-square error loss and perceptual loss preserves many structural details in the CT image. This objective function does not suffer from over smoothing and blurring effects causing by per-pixel loss and grid-like artifacts resulting from perceptual loss. The experiments show that each modification to the network improves the outcome while changing the complexity of the network, minimally.

Motion-compensated data decomposition algorithm to accelerate dynamic cardiac MRI.

OBJECTIVES: In dynamic cardiac magnetic resonance imaging (MRI), the spatiotemporal resolution is often limited by low imaging speed. Compressed sensing (CS) theory can be applied to improve imaging speed and spatiotemporal resolution. The combination of compressed sensing and low-rank matrix completion represents an attractive means to further increase imaging speed. By extending prior work, a Motion-Compensated Data Decomposition (MCDD) algorithm is proposed to improve the performance of CS for accelerated dynamic cardiac MRI. MATERIALS AND METHODS: The process of MCDD can be described as follows: first, we decompose the dynamic images into a low-rank (L) and a sparse component (S). The L component includes periodic motion in the background, since it is highly correlated among frames, and the S component corresponds to respiratory motion. A motion-estimation/motion-compensation (ME-MC) algorithm is then applied to the low-rank component to reconstruct a cardiac motion compensated dynamic cardiac MRI. RESULTS: With validations on the numerical phantom and in vivo cardiac MRI data, we demonstrate the utility of the proposed scheme in significantly improving compressed sensing reconstructions by minimizing motion artifacts. The proposed method achieves higher PSNR and lower MSE and HFEN for medium to high acceleration factors. CONCLUSION: The proposed method is observed to yield reconstructions with minimal spatiotemporal blurring and motion artifacts in comparison to the existing state-of-the-art methods.

Pure Vision Transformer (CT-ViT) with Noise2Neighbors Interpolation for Low-Dose CT Image Denoising.

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

Dilated Convolution ResNet with Boosting Attention Modules and Combined Loss Functions for LDCT Image Denoising.

With the increasing concern regarding the radiation exposure of patients undergoing computed tomography (CT) scans, researchers have been using deep learning techniques to improve the quality of denoised low-dose CT (LDCT) images. In this paper, a cascaded dilated residual network (ResNet) with integrated attention modules, specifically spatial- and channel- attention modules, is proposed. This experiment demonstrated how these attention modules improved the denoised CT image by testing a simple ResNet with and without the modules. Further, an investigation regarding the effectiveness of per-pixel loss, perceptual loss via VGG16-Net, and structural dissimilarity loss functions is also covered through an ablation experiment. By knowing how these loss functions affect the output denoised images, a combination of the these loss function is then proposed which aims to prevent edge over-smoothing, enhance textural details and finally, preserve structural details on the denoised images. Finally, a bench testing was also done by comparing the visual and quantitative results of the proposed model with the state-of-the-art models such as block matching 3D (BM3D), patch-GAN and dilated convolution with edge detection layer (DRL-E-MP) for accuracy.

Gradient-based Optimization Algorithm for Hybrid Loss Function in Low-dose CT Denoising.

Deep learning techniques have emerged in de-noising low-dose computed tomography (CT) images to avoid the potential health risks of high ionizing radiation dose on patients. Although these post-processing methods display high quality denoised images, the denoising performance still has the potential to improve. The primary purpose of this work was to determine and analyze the most effective and efficient hybrid loss function in deep learning (DL)-based denoising network. Objective functions in deep learning algorithms are the main keys for optimizing the parameters of a network and can affect the quality of the denoised image significantly. Hence, this work examined the various combinations of the most common objective functions in CT denoising networks, namely L1 loss, per-pixel loss, perceptual loss, and structural dissimilarity loss. Further, a hyperparameter learning algorithm was also introduced to find the best scalable factors of the loss functions in each hybrid loss function combination. For simplic-ity, RED-CNN was used in this study to easily demonstrate the performance of the losses during the denoising process. Based on this experiment, the balance between these loss function via the gradient-based optimization algorithm could help in the generalizability prediction of designing future CT denoising networks.

Low Dose CT Image Denoising Using Boosting Attention Fusion GAN with Perceptual Loss.

Image denoising of Low-dose computed tomography (LDCT) images has continues to receive attention in the research community due to ongoing concerns about high-dose radiation exposure of patients for diagnosis. The use of low radiation CT image, however, could lead to inaccurate diagnosis due to the presence of noise. Deep learning techniques are being integrated into denoising methods to address this problem. In this paper, a General Adversarial Network (GAN) composed of boosting fusion of spatial and channel attention modules is proposed. These modules are embedded in the denoiser to address the limitations of other GAN-based denoising models that tend to only focus on the local processing and neglect the dependencies of creating feature maps with spatial- and channel- wise image characteristics. This study aims to preserve structural details of LDCT images by applying boosting attention modules, prevents edge over-smoothing by integrating perceptual loss via VGG16 pre-trained network, and finally, improves the computational efficiency by taking advantage of deep learning techniques and GPU parallel computation.

Cascaded Learning with Generative Adversarial Networks for Low Dose CT Denoising.

CT machines can be tuned in order to reduce the radiation dose used for imaging, yet reducing the radiation dose results in noisy images which are not suitable in clinical practice. In order for low dose CT to be used effectively in practice this issue must be addressed. Generative Adversarial Networks (GAN) have been used widely in computer vision research and have proven themselves as a powerful tool for producing images with high perceptual quality. In this work we use a cascade of two neural networks, the first is a Generative Adversarial Network and the second is a Deep Convolutional Neural Network. The first network generates a denoised sample which is then fine-tuned by the second network via residue learning. We show that our cascaded method outperforms related works and more effectively reconstructs fine structural details in low contrast regions of the image.

Empirical Mode Decomposition Based Hyperspectral Data Analysis for Brain Tumor Classification.

The capability of Hyperspectral Imaging (HSI) in rapidly acquiring abundant reflectance data in a non-invasive manner, makes it an ideal tool for obtaining diagnostic information about tissue pathology. Identifying wavelengths that provide the most discriminatory clues for specific pathologies will greatly assist in understanding their underlying biochemical characteristics. In this paper, we propose an efficient and computationally inexpensive method for determining the most relevant spectral bands for brain tumor classification. Empirical mode decomposition was used in combination with extrema analysis to extract the relevant bands based on the morphological characteristics of the spectra. The results of our experiments indicate that the proposed method outperforms the benchmark in reducing computational complexity while performing comparably with a 7-times reduction in the feature-set for classification on the test data.

Use of dynamic CT in acute respiratory distress syndrome (ARDS) with comparison of positive and negative pressure ventilation.

Negative pressure ventilation via an external device ('iron lung') has the potential to provide better oxygenation with reduced barotrauma in patients with ARDS. This study was designed to see if oxygenation differences between positive and negative ventilation could be explained by CT. Six anaesthetized rabbits had ARDS induced by repeated saline lavage. Rabbits were ventilated with positive pressure ventilation (PPV) and negative pressure ventilation (NPV) in turn. Dynamic CT images were acquired over the respiratory cycle. A computer-aided method was used to segment the lung and calculate the range of CT densities within each slice. Volumes of ventilated lung and atelectatic lung were measured over the respiratory cycle. NPV was associated with an increased percentage of ventilated lung and decreased percentage of atelectatic lung. The most significant differences in ventilation and atelectasis were seen at mid-inspiration and mid-expiration (ventilated lung NPV = 61%, ventilated lung PPV = 47%, p < 0.001; atelectatic lung NPV = 10%, atelectatic lung PPV 19%, p < 0.001). Aeration differences were not significant at end-inspiration. Dynamic CT can show differences in lung aeration between positive and negative ventilation in ARDS. These differences would not be appreciated if only static breath-hold CT was used.

Low-dose CT Denoising Using Edge Detection Layer and Perceptual Loss.

Low-dose CT imaging is a valid approach to reduce patients' exposure to X-ray radiation. However, reducing X-ray current increases noise and artifacts in the reconstructed CT images. Deep neural networks have been successfully employed to remove noise from low-dose CT images. This study proposes two novel techniques to boost the performance of a neural network with minimal change in the complexity. First, a non-trainable edge detection layer is proposed that extracts four edge maps from the input image. The layer improves quantitative metrics (PSNR and SSIM) and helps to predict a CT image with more precise boundaries. Next, a joint function of mean-square error and perceptual loss is employed to optimize the network. Using the perceptual loss helps to preserve structural detail; however, it adds check-board artifacts to the output. The proposed joint objective function takes advantage of the benefits offered by each loss. It improves the over-smoothing problem caused by mean-square error and the check-board artifacts caused by perceptual loss.

Nonrigid motion compensation in compressed sensing reconstruction of cardiac cine MRI.

In this work, a robust nonrigid motion compensation approach, is applied to the compressed sensing reconstruction of dynamic cardiac cine MRI sequences. Respiratory and cardiac motion separation coupled with a registration algorithm is used to provide accurate reconstruction of dynamic cardiac images. The proposed scheme employs a variable splitting based optimization strategy to enable joint motion estimation along with reconstruction. We define the recovery as an energy minimization scheme utilizing an objective function that combines data consistency, spatial smoothness, and motion penalties. The validation of the proposed algorithm using numerical phantom and in-vivo cine MRI data demonstrates reconstruction of cardiac MRI data with less spatio-temporal blurring and motion artifacts from extensively under-sampled data. The proposed method is observed to provide improved reconstructions over state-of-the-art motion compensation schemes.

Low-dose CT Denoising with Dilated Residual Network.

Low-dose Computed Tomography (CT) is considered a solution for reducing the risk of X-ray radiation; however, lowering the X-ray current results in a degraded reconstructed image. To improve the quality of the image, different noise removal techniques have been proposed. Con- volutional neural networks also have shown promising results in denoising the low-dose CT images. In this paper, a deep residual network with dilated convolution is proposed. The identity mappings pass the signal to the higher layers and improve the performance of the network and its training time. Moreover, employing dilated convolution helps to increase the receptive field faster. Dilated convolution makes it possible to achieve good results with fewer layers and less computational costs. The proposed network learns end to end mapping from low-dose to normal-dose CT images.

Transesophageal 2D ultrasound to 3D computed tomography registration for the guidance of a cardiac arrhythmia therapy.

The work aims to develop a new image-processing method to improve the guidance of transesophageal high intensity focused ultrasound (HIFU) atrial fibrillation therapy. Our proposal is a novel registration approach that aligns intraoperative 2D ultrasound with preoperative 3D-CT information. This approach takes advantage of the anatomical constraints imposed at the transesophageal HIFU probe to simplify the registration process. Our proposed method has been evaluated on a physical phantom and on real clinical data.

Colony size measurement of the yeast gene deletion strains for functional genomics.

BACKGROUND: Numerous functional genomics approaches have been developed to study the model organism yeast, Saccharomyces cerevisiae, with the aim of systematically understanding the biology of the cell. Some of these techniques are based on yeast growth differences under different conditions, such as those generated by gene mutations, chemicals or both. Manual inspection of the yeast colonies that are grown under different conditions is often used as a method to detect such growth differences. RESULTS: Here, we developed a computerized image analysis system called Growth Detector (GD), to automatically acquire quantitative and comparative information for yeast colony growth. GD offers great convenience and accuracy over the currently used manual growth measurement method. It distinguishes true yeast colonies in a digital image and provides an accurate coordinate oriented map of the colony areas. Some post-processing calculations are also conducted. Using GD, we successfully detected a genetic linkage between the molecular activity of the plant-derived antifungal compound berberine and gene expression components, among other cellular processes. A novel association for the yeast mek1 gene with DNA damage repair was also identified by GD and confirmed by a plasmid repair assay. The results demonstrate the usefulness of GD for yeast functional genomics research. CONCLUSION: GD offers significant improvement over the manual inspection method to detect relative yeast colony size differences. The speed and accuracy associated with GD makes it an ideal choice for large-scale functional genomics investigations.

A dictionary learning approach for human sperm heads classification.

BACKGROUND AND OBJECTIVE: To diagnose infertility in men, semen analysis is conducted in which sperm morphology is one of the factors that are evaluated. Since manual assessment of sperm morphology is time-consuming and subjective, automatic classification methods are being developed. Automatic classification of sperm heads is a complicated task due to the intra-class differences and inter-class similarities of class objects. In this research, a Dictionary Learning (DL) technique is utilized to construct a dictionary of sperm head shapes. This dictionary is used to classify the sperm heads into four different classes. METHODS: Square patches are extracted from the sperm head images. Columnized patches from each class of sperm are used to learn class-specific dictionaries. The patches from a test image are reconstructed using each class-specific dictionary and the overall reconstruction error for each class is used to select the best matching class. Average accuracy, precision, recall, and F-score are used to evaluate the classification method. The method is evaluated using two publicly available datasets of human sperm head shapes. RESULTS: The proposed DL based method achieved an average accuracy of 92.2% on the HuSHeM dataset, and an average recall of 62% on the SCIAN-MorphoSpermGS dataset. The results show a significant improvement compared to a previously published shape-feature-based method. We have achieved high-performance results. In addition, our proposed approach offers a more balanced classifier in which all four classes are recognized with high precision and recall. CONCLUSIONS: In this paper, we use a Dictionary Learning approach in classifying human sperm heads. It is shown that the Dictionary Learning method is far more effective in classifying human sperm heads than classifiers using shape-based features. Also, a dataset of human sperm head shapes is introduced to facilitate future research.

Shear wave elastography using Wigner-Ville distribution: a simulated multilayer media study.

Shear Wave Elastography (SWE) is a quantitative ultrasound-based imaging modality for distinguishing normal and abnormal tissue types by estimating the local viscoelastic properties of the tissue. These properties have been estimated in many studies by propagating ultrasound shear wave within the tissue and estimating parameters such as speed of wave. Vast majority of the proposed techniques are based on the cross-correlation of consecutive ultrasound images. In this study, we propose a new method of wave detection based on time-frequency (TF) analysis of the ultrasound signal. The proposed method is a modified version of the Wigner-Ville Distribution (WVD) technique. The TF components of the wave are detected in a propagating ultrasound wave within a simulated multilayer tissue and the local properties are estimated based on the detected waves. Image processing techniques such as Alternative Sequential Filters (ASF) and Circular Hough Transform (CHT) have been utilized to improve the estimation of TF components. This method has been applied to a simulated data from Wave3000™ software (CyberLogic Inc., New York, NY). This data simulates the propagation of an acoustic radiation force impulse within a two-layer tissue with slightly different viscoelastic properties between the layers. By analyzing the local TF components of the wave, we estimate the longitudinal and shear elasticities and viscosities of the media. This work shows that our proposed method is capable of distinguishing between different layers of a tissue.

Compressed sensing reconstruction of cardiac cine MRI using golden angle spiral trajectories.

In dynamic cardiac cine Magnetic Resonance Imaging (MRI), the spatiotemporal resolution is limited by the low imaging speed. Compressed sensing (CS) theory has been applied to improve the imaging speed and thus the spatiotemporal resolution. The purpose of this paper is to improve CS reconstruction of under sampled data by exploiting spatiotemporal sparsity and efficient spiral trajectories. We extend k-t sparse algorithm to spiral trajectories to achieve high spatio temporal resolutions in cardiac cine imaging. We have exploited spatiotemporal sparsity of cardiac cine MRI by applying a 2D+time wavelet-Fourier transform. For efficient coverage of k-space, we have used a modified version of multi shot (interleaved) spirals trajectories. In order to reduce incoherent aliasing artifact, we use different random undersampling pattern for each temporal frame. Finally, we have used nonuniform fast Fourier transform (NUFFT) algorithm to reconstruct the image from the non-uniformly acquired samples. The proposed approach was tested in simulated and cardiac cine MRI data. Results show that higher acceleration factors with improved image quality can be obtained with the proposed approach in comparison to the existing state-of-the-art method. The flexibility of the introduced method should allow it to be used not only for the challenging case of cardiac imaging, but also for other patient motion where the patient moves or breathes during acquisition.

Estimation of intraoperative brain shift by combination of stereovision and doppler ultrasound: phantom and animal model study.

PURPOSE: Combination of various intraoperative imaging modalities potentially can reduce error of brain shift estimation during neurosurgical operations. In the present work, a new combination of surface imaging and Doppler US images is proposed to calculate the displacements of cortical surface and deformation of internal vessels in order to estimate the targeted brain shift using a Finite Element Model (FEM). Registration error in each step and the overall performance of the method are evaluated. METHODS: The preoperative steps include constructing a FEM from MR images and extracting vascular tree from MR Angiography (MRA). As the first intraoperative step, after the craniotomy and with the dura opened, a designed checkerboard pattern is projected on the cortex surface and projected landmarks are scanned and captured by a stereo camera (Int J Imaging Syst Technol 23(4):294-303, 2013. doi: 10.1002/ima.22064 ). This 3D point cloud should be registered to boundary nodes of FEM in the region of interest. For this purpose, we developed a new non-rigid registration method, called finite element drift that is more compatible with the underlying nature of deformed object. The presented algorithm outperforms other methods such as coherent point drift when the deformation is local or non-coherent. After registration, the acquired displacement vectors are used as boundary conditions for FE model. As the second step, by tracking a 2D Doppler ultrasound probe swept on the parenchyma, a 3D image of deformed vascular tree is constructed. Elastic registration of this vascular point cloud to the corresponding preoperative data results the second series of displacement vector applicable to closest internal nodes of FEM. After running FE analysis, the displacement of all nodes is calculated. The brain shift is then estimated as displacement of nodes in boundary of a deep target, e.g., a tumor. We used intraoperative MR (iMR) images as the references for measuring the performance of the brain shift estimator. In the present study, two set of tests were performed using: (a) a deformable brain phantom with surface data and (b) an alive brain of an approximately big dog with surface data and US Doppler images. In our designed phantom, small tubes connected to an inflatable balloon were considered as displaceable targets and in the animal model, the target was modeled by a cyst which was created by an injection. RESULTS: In the phantom study, the registration error for the surface points before FE analysis and for the target points after running FE model were <0.76 and 1.4 mm, respectively. In a real condition of operating room for animal model, the registration error was about 1 mm for the surface, 1.9 mm for the vascular tree and 1.55 mm for the target points. CONCLUSIONS: The proposed projected surface imaging in conjunction with the Doppler US data combined in a powerful biomechanical model can result an acceptable performance in calculation of deformation during surgical navigation. However, the projected landmark method is sensitive to ambient light and surface conditions and the Doppler ultrasound suffers from noise and 3D image construction problems, the combination of these two methods applied on a FEM has an eligible performance.

Multi dose computed tomography image fusion based on hybrid sparse methodology.

With the increasing utilization of X-ray Computed Tomography (CT) in medical diagnosis, obtaining higher quality image with lower exposure to radiation has become a highly challenging task in image processing. In this paper, a novel sparse fusion algorithm is proposed to address the problem of lower Signal to Noise Ratio (SNR) in low dose CT images. Initial fused image is obtained by combining low dose and medium dose images in sparse domain, utilizing the Dual Tree Complex Wavelet Transform (DTCWT) dictionary which is trained by high dose image. And then, the strongly focused image is obtained by determining the pixels of source images which have high similarity with the pixels of the initial fused image. Final denoised image is obtained by fusing strongly focused image and decomposed sparse vectors of source images, thereby preserving the edges and other critical information needed for diagnosis. This paper demonstrates the effectiveness of the proposed algorithm both quantitatively and qualitatively.

Low-dose computed tomography image denoising based on joint wavelet and sparse representation.

Image denoising and signal enhancement are the most challenging issues in low dose computed tomography (CT) imaging. Sparse representational methods have shown initial promise for these applications. In this work we present a wavelet based sparse representation denoising technique utilizing dictionary learning and clustering. By using wavelets we extract the most suitable features in the images to obtain accurate dictionary atoms for the denoising algorithm. To achieve improved results we also lower the number of clusters which reduces computational complexity. In addition, a single image noise level estimation is developed to update the cluster centers in higher PSNRs. Our results along with the computational efficiency of the proposed algorithm clearly demonstrates the improvement of the proposed algorithm over other clustering based sparse representation (CSR) and K-SVD methods.