Convolutional neural networks (CNNs) have been widely used for low-dose CT (LDCT) image denoising. To alleviate the over-smoothing effect caused by conventional mean squared error, researchers often resort to adversarial training to achieve faithful structural and texture recovery. On one hand, such adversarial training is typically difficult to train and may lead to a potential CT value shift. On the other hand, these CNNs-based denoising models usually generalize poorly to new unseen dose levels. Recently, diffusion models have exhibited higher image quality and stable trainability compared to other generative models. Therefore, we present a Contextual Conditional Diffusion model (CoCoDiff) for low-dose CT denoising, which aims to address the issues of existing denoising models. More specifically, during the training stage, we train a noise estimation network to gradually convert a residual image to a Gaussian distribution based on a Markov chain with a low-dose image as the condition. During the inference stage, the Markov chain is reversed to generate the residual image from random Gaussian noise. In addition, the contextual information of adjacent slices is also utilized for noise estimation to suppress potential structural distortion. Experimental results on Mayo LDCT datasets show that the proposed model can restore faithful structural details and generalize well to other unseen noise levels.
Stimulated emission depletion (STED), as one of the emerging super-resolution techniques, defines a state-of-the-art image resolution method. It has been developed into a universal fluorescent imaging tool over the past several years. The currently best available lateral resolution offered by STED is around 20 nm, but in real live cell imaging applications, the regular resolution offered through this mechanism is around 100 nm, limited by phototoxicity. Many critical biological structures are below this resolution level. Hence, it will be invaluable to improve the STED resolution through postprocessing techniques. We propose a deep adversarial network for improving the STED resolution significantly, which takes an STED image as an input, relies on physical modeling to obtain training data, and outputs a “self-refined” counterpart image at a higher resolution level. In other words, we use the prior knowledge on the STED point spread function and the structural information about the cells to generate simulated labeled data pairs for network training. Our results suggest that 30-nm resolution can be achieved from a 60-nm resolution STED image, and in our simulation and experiments, the structural similarity index values between the label and output result reached around 0.98, significantly higher than those obtained using the Lucy–Richardson deconvolution method and a state-of-the-art UNet-based super-resolution network.
Recently, deep learning has become the mainstream method in multiple fields of artificial intelligence / machine learning (AI/ML) applications, including medical imaging. Encouraged by the neural diversity in the human body, our group proposed to replace the inner product in the current artificial neuron with a quadratic operation on inputs (called quadratic neuron) for deep learning. Since the representation capability at the cellular level is enhanced by the quadratic neuron, we are motivated to build network architectures and evaluate the potential of quadratic neurons towards “quadratic deep learning”. Along this direction, our previous theoretical studies have shown advantages of quadratic neurons and quadratic networks in terms of efficiency and representation. In this paper, we prototype a quadratic residual neural network (Q-ResNet) by incorporating quadratic neurons into a convolutional residual structure, and then deploy it for CT metal artifact reduction. Also, we report our experiments on a simulated dataset to show that Q-ResNet performs better than the classic NMAR algorithm.
Tomographically measuring the temperature distribution inside a human body has important and immediate clinical applications - including thermal ablative and hyperthermic treatment of cancers, and will enable novel solutions such as opening the blood-brain barrier for therapeutics and robotic surgery in the future. A high intensity focused ultrasound (HIFU) device can heat tumor tissues to 50-90 °C locally within seconds. Thus, accurate, real-time control of the thermal dose is critical to eliminate tumors while minimizing damage to surrounding healthy tissue. This study investigates the feasibility of using deep learning to improve the accuracy of low-dose CT (LDCT) thermometry. CT thermometry relies on thermal expansion coefficients, which is prone to inter-patient variability, and is also affected by image noise and artifacts. To demonstrate that deep neural networks can compensate for these factors, 1,000 computer-generated CT phantoms with simulated heating spots were used in training both a “divide and conquer” and “end to end” approach. In the first strategy, the first encoder-decoder network corrected for beam hardening and Poisson noise in the image domain, while a second fine-tuned differences between predicted and ground truth heat maps. The second strategy is identical to the first, except only a single convolutional autoencoder was used as the CT images were not pre-cleaned. Ultimately, the two-part divide and conquer network increased thermal accuracy substantially, demonstrating exciting future potential for the use of deep learning in this field.
X-ray computed tomography (CT) reconstructs cross-sectional images from projection data. However, ionizing X-ray radiation associated with CT scanning might induce cancer and genetic damage and raises public concerns. Therefore, the reduction of radiation dose has attracted major attention. Few-view CT image reconstruction is an important topic to reduce the radiation dose. Recently, data-driven algorithms have shown great potential to solve the few-view CT problem. In this paper, we develop a dual network architecture (DNA) for reconstructing images directly from sinograms. In the proposed DNA method, a point-wise fully-connected layer learns the backprojection process requesting significantly less memory than the prior art and with O(C×N×NC) parameters where N and Nc denote the dimension of reconstructed images and number of projections respectively. C is an adjustable parameter that can be set as low as 1. Our experimental results demonstrate that DNA produces a competitive performance over the other state-of-the-art methods.Interestingly, natural images can be used to pre-train DNA to avoid overfitting when the amount of real patient images is limited.
Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) are widely used for screening, diagnosis and imageguided therapeutics. Due to physical, technical and economical limitations, it is impossible for MRI and CT scanners to target ideal image resolution. Given the nominal imaging performance, how to improve image resolution has been a hot topic, and referred to as super-resolution research. As a promising method for super-resolution, over recent years deep learning has shown a great potential especially in deblurring natural images. In this paper, based on the neural network model termed as GAN-CIRCLE (Constrained by the Identical, Residual, Cycle Learning Ensemble), we adapt this neural network for achieving super-resolution for both MRI and CT. In this study, we demonstrate two-fold resolution enhancement for MRI and CT with the same network architecture.
Cone-beam breast computed tomography (CT) provides true 3D breast images with isotropic resolution and highcontrast information, detecting calcifications as small as a few hundred microns and revealing subtle tissue differences. However, breast is highly sensitive to x-ray radiation. It is critically important for healthcare to reduce radiation dose. Few-view cone-beam CT only uses a fraction of x-ray projection data acquired by standard cone-beam breast CT, enabling significant reduction of the radiation dose. However, insufficient sampling data would cause severe streak artifacts in images reconstructed using conventional methods. We propose a deep-learning-based method for the image reconstruction to establish a residual neural network model, which is applied for few-view breast CT to produce high quality breast CT images. In this study, we respectively evaluate the breast image reconstruction from one third and one quarter of x-ray projection views of the standard cone-beam breast CT. Based on clinical breast imaging dataset, we perform a supervised learning to train the neural network from few-view CT images to corresponding full-view CT images. Experimental results show that the deep learning-based image reconstruction method allows few-view breast CT to achieve a radiation dose <6mGy per cone-beam CT scan which is a threshold set by FDA for mammographic screening.
This paper introduces a generative adversarial network (GAN) for low-dose CT (LDCT) simulation, which is an inverse process for network-based low-dose CT denoising. Within our GAN framework, the generator is an encoder-decoder network with a shortcut connection to produce realistic noisy LDCT images. To ensure satisfactory results, a conditional batch normalization layer is incorporated into the bottleneck between the encoder and the decoder. After the model is trained, a Gaussian noise generator serves as the latent variable controlling the noise in generated CT images. With the Mayo Low-dose CT Challenge dataset, the proposed network was trained on image patches, and then produced full-size low-dose CT images of different noise distributions at various noise levels. The network-generated low-dose CT images can be used to test the robustness of the current low-dose CT denoising models and also help perform other imaging tasks such as optimization of radiation dose to patients and evaluation of model observers.
Photoacoustic tomography seeks to reconstruct an acoustic initial pressure distribution from the measurement of the ultrasound waveforms. Conventional methods assume a-prior knowledge of the sound speed distribution, which practically is unknown. One way to circumvent the issue is to simultaneously reconstruct both the acoustic initial pressure and speed. In this article, we develop a novel data-driven method that integrates an advanced deep neural network through model-based iteration. The image of the initial pressure is significantly improved in our numerical simulation.
Recently, deep learning has transformed many fields including medical imaging. Inspired by diversity of biological neurons, our group proposed quadratic neurons in which the inner product in current artificial neurons is replaced with a quadratic operation on inputs, thereby enhancing the capability of an individual neuron. Along this direction, we are motivated to evaluate the power of quadratic neurons in representative network architectures, towards “quadratic neuron based deep learning”. In this regard, our prior theoretical studies have shown important merits of quadratic neurons and networks. In this paper, we use quadratic neurons to construct an encoder-decoder structure, referred to as the quadratic autoencoder, and apply it for low-dose CT denoising. Then, we perform experiments on the Mayo low-dose CT dataset to demonstrate that the quadratic autoencoder yields a better denoising performance.
Over the past few years, deep neural networks have made significant processes in denoising low-dose CT images. A trained denoising network, however, may not generalize very well to different dose levels, which follows from the dose-dependent noise distribution. To address this practically, a trained network requires re-training to be applied to a new dose level, which limits the generalization abilities of deep neural networks for clinical applications. This article introduces a deep learning approach that does not require re-training and relies on a transfer learning strategy. More precisely, the transfer learning framework utilizes a progressive denoising model, where an elementary neural network serves as a basic denoising unit. The basic units are then cascaded to successively process towards a denoising task; i.e. the output of one network unit is the input to the next basic unit. The denoised image is then a linear combination of outputs of the individual network units. To demonstrate the application of this transfer learning approach, a basic CNN unit is trained using the Mayo low- dose CT dataset. Then, the linear parameters of the successive denoising units are trained using a different image dataset, i.e. the MGH low-dose CT dataset, containing CT images that were acquired at four different dose levels. Compared to a commercial iterative reconstruction approach, the transfer learning framework produced a substantially better denoising performance.
Artifacts resulting from metal objects have been a persistent problem in CT images over the last four decades. A common
approach to overcome their effects is to replace corrupt projection data with values synthesized from an interpolation
scheme or by reprojection of a prior image. State-of-the-art correction methods, such as the interpolation- and
normalization-based algorithm NMAR, often do not produce clinically satisfactory results. Residual image artifacts remain
in challenging cases and even new artifacts can be introduced by the interpolation scheme. Metal artifacts continue to be
a major impediment, particularly in radiation and proton therapy planning as well as orthopedic imaging. A new solution
to the long-standing metal artifact reduction (MAR) problem is deep learning, which has been successfully applied to
medical image processing and analysis tasks. In this study, we combine a convolutional neural network (CNN) with the
state-of-the-art NMAR algorithm to reduce metal streaks in critical image regions. Training data was synthesized from CT
simulation scans of a phantom derived from real patient images. The CNN is able to map metal-corrupted images to
artifact-free monoenergetic images to achieve additional correction on top of NMAR for improved image quality. Our
results indicate that deep learning is a novel tool to address CT reconstruction challenges, and may enable more accurate
tumor volume estimation for radiation therapy planning.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.