Learned Primal-Dual Reconstruction

Adapting model-based deep learning to multiple acquisition conditions: Ada-MoDL

PURPOSE

The aim of this work is to introduce a single model-based deep network that can provide high-quality reconstructions from undersampled parallel MRI data acquired with multiple sequences, acquisition settings, and field strengths.

METHODS

A single unrolled architecture, which offers good reconstructions for multiple acquisition settings, is introduced. The proposed scheme adapts the model to each setting by scaling the convolutional neural network (CNN) features and the regularization parameter with appropriate weights. The scaling weights and regularization parameter are derived using a multilayer perceptron model from conditional vectors, which represents the specific acquisition setting. The perceptron parameters and the CNN weights are jointly trained using data from multiple acquisition settings, including differences in field strengths, acceleration, and contrasts. The conditional network is validated using datasets acquired with different acquisition settings.

RESULTS

The comparison of the adaptive framework, which trains a single model using the data from all the settings, shows that it can offer consistently improved performance for each acquisition condition. The comparison of the proposed scheme with networks that are trained independently for each acquisition setting shows that it requires less training data per acquisition setting to offer good performance.

CONCLUSION

The Ada-MoDL framework enables the use of a single model-based unrolled network for multiple acquisition settings. In addition to eliminating the need to train and store multiple networks for different acquisition settings, this approach reduces the training data needed for each acquisition setting.

AI for PET image reconstruction

Image reconstruction for positron emission tomography (PET) has been developed over many decades, with advances coming from improved modelling of the data statistics and improved modelling of the imaging physics. However, high noise and limited spatial resolution have remained issues in PET imaging, and state-of-the-art PET reconstruction has started to exploit other medical imaging modalities (such as MRI) to assist in noise reduction and enhancement of PET's spatial resolution. Nonetheless, there is an ongoing drive towards not only improving image quality, but also reducing the injected radiation dose and reducing scanning times. While the arrival of new PET scanners (such as total body PET) is helping, there is always a need to improve reconstructed image quality due to the time and count limited imaging conditions. Artificial intelligence (AI) methods are now at the frontier of research for PET image reconstruction. While AI can learn the imaging physics as well as the noise in the data (when given sufficient examples), one of the most common uses of AI arises from exploiting databases of high-quality reference examples, to provide advanced noise compensation and resolution recovery. There are three main AI reconstruction approaches: (i) direct data-driven AI methods which rely on supervised learning from reference data, (ii) iterative (unrolled) methods which combine our physics and statistical models with AI learning from data, and (iii) methods which exploit AI with our known models, but crucially can offer benefits even in the absence of any example training data whatsoever. This article reviews these methods, considering opportunities and challenges of AI for PET reconstruction.

Physics-driven Learned Deconvolution of Multi-spectral Cellular MRI with Radial Sampling

In fluorine-19 (19F) cellular MRI, detection of multiple cell targets requires the ability to unmix images corresponding to different tracer molecules with different chemical shifts. The resulting chemical shift artifacts of conventional Cartesian sampling are well-defined, appearing as 'ghost images' along the readout direction. However, a key challenge with radial sampling is that frequency offsets lead to non-linear smearing artifacts throughout the image. Thus, proper modeling of forward sensing operators is crucial, as the successful deconvolution of artifacts relies on the joint design of acquisition scheme and sampling pattern. In this work, we aim to address these perspectives through the lens of radial spectral deconvolution. Our goal is to develop suitable modeling of the radial chemical shift artifacts using Radon transform, which will guide our design of sensing operators that favor specific multi-spectral imaging tasks. To effectively unmix the component images under low SNR regime, we will further exploit physics-informed learning-based unrolling strategies that enable simultaneous artifact removal and weak signal detection, both of particular interest in 19F MRI. particular interest in 19F MRI.

CAIR: Combining integrated attention with iterative optimization learning for sparse-view CT reconstruction

Sparse-view CT is an efficient way for low dose scanning but degrades image quality. Inspired by the successful use of non-local attention in natural image denoising and compression artifact removal, we proposed a network combining integrated attention and iterative optimization learning for sparse-view CT reconstruction (CAIR). Specifically, we first unrolled the proximal gradient descent into a deep network and added an enhanced initializer between the gradient term and the approximation term. It can enhance the information flow between different layers, fully preserve the image details, and improve the network convergence speed. Secondly, the integrated attention module was introduced into the reconstruction process as a regularization term. It adaptively fuses the local and non-local features of the image which are used to reconstruct the complex texture and repetitive details of the image, respectively. Note that we innovatively designed a one-shot iteration strategy to simplify the network structure and reduce the reconstruction time while maintaining image quality. Experiments showed that the proposed method is very robust and outperforms state-of-the-art methods in terms of both quantitative and qualitative, greatly improving the preservation of structures and the removal of artifacts.

Sinogram upsampling using Primal-Dual UNet for undersampled CT and radial MRI reconstruction

Computed tomography (CT) and magnetic resonance imaging (MRI) are two widely used clinical imaging modalities for non-invasive diagnosis. However, both of these modalities come with certain problems. CT uses harmful ionising radiation, and MRI suffers from slow acquisition speed. Both problems can be tackled by undersampling, such as sparse sampling. However, such undersampled data leads to lower resolution and introduces artefacts. Several techniques, including deep learning based methods, have been proposed to reconstruct such data. However, the undersampled reconstruction problem for these two modalities was always considered as two different problems and tackled separately by different research works. This paper proposes a unified solution for both sparse CT and undersampled radial MRI reconstruction, achieved by applying Fourier transform-based pre-processing on the radial MRI and then finally reconstructing both modalities using sinogram upsampling combined with filtered back-projection. The Primal-Dual network is a deep learning based method for reconstructing sparsely-sampled CT data. This paper introduces Primal-Dual UNet, which improves the Primal-Dual network in terms of accuracy and reconstruction speed. The proposed method resulted in an average SSIM of 0.932±0.021 while performing sparse CT reconstruction for fan-beam geometry with a sparsity level of 16, achieving a statistically significant improvement over the previous model, which resulted in 0.919±0.016. Furthermore, the proposed model resulted in 0.903±0.019 and 0.957±0.023 average SSIM while reconstructing undersampled brain and abdominal MRI data with an acceleration factor of 16, respectively - statistically significant improvements over the original model, which resulted in 0.867±0.025 and 0.949±0.025. Finally, this paper shows that the proposed network not only improves the overall image quality, but also improves the image quality for the regions-of-interest: liver, kidneys, and spleen; as well as generalises better than the baselines in presence the of a needle.

Integrating Artificial Intelligence Tools in the Clinical Research Setting: The Ovarian Cancer Use Case

Artificial intelligence (AI) methods applied to healthcare problems have shown enormous potential to alleviate the burden of health services worldwide and to improve the accuracy and reproducibility of predictions. In particular, developments in computer vision are creating a paradigm shift in the analysis of radiological images, where AI tools are already capable of automatically detecting and precisely delineating tumours. However, such tools are generally developed in technical departments that continue to be siloed from where the real benefit would be achieved with their usage. Significant effort still needs to be made to make these advancements available, first in academic clinical research and ultimately in the clinical setting. In this paper, we demonstrate a prototype pipeline based entirely on open-source software and free of cost to bridge this gap, simplifying the integration of tools and models developed within the AI community into the clinical research setting, ensuring an accessible platform with visualisation applications that allow end-users such as radiologists to view and interact with the outcome of these AI tools.

DeepSTI: Towards tensor reconstruction using fewer orientations in susceptibility tensor imaging

Susceptibility tensor imaging (STI) is an emerging magnetic resonance imaging technique that characterizes the anisotropic tissue magnetic susceptibility with a second-order tensor model. STI has the potential to provide information for both the reconstruction of white matter fiber pathways and detection of myelin changes in the brain at mm resolution or less, which would be of great value for understanding brain structure and function in healthy and diseased brain. However, the application of STI in vivo has been hindered by its cumbersome and time-consuming acquisition requirement of measuring susceptibility induced MR phase changes at multiple head orientations. Usually, sampling at more than six orientations is required to obtain sufficient information for the ill-posed STI dipole inversion. This complexity is enhanced by the limitation in head rotation angles due to physical constraints of the head coil. As a result, STI has not yet been widely applied in human studies in vivo. In this work, we tackle these issues by proposing an image reconstruction algorithm for STI that leverages data-driven priors. Our method, called DeepSTI, learns the data prior implicitly via a deep neural network that approximates the proximal operator of a regularizer function for STI. The dipole inversion problem is then solved iteratively using the learned proximal network. Experimental results using both simulation and in vivo human data demonstrate great improvement over state-of-the-art algorithms in terms of the reconstructed tensor image, principal eigenvector maps and tractography results, while allowing for tensor reconstruction with MR phase measured at much less than six different orientations. Notably, promising reconstruction results are achieved by our method from only one orientation in human in vivo, and we demonstrate a potential application of this technique for estimating lesion susceptibility anisotropy in patients with multiple sclerosis.

Similarity-Driven Fine-Tuning Methods for Regularization Parameter Optimization in PET Image Reconstruction

We present an adaptive method for fine-tuning hyperparameters in edge-preserving regularization for PET image reconstruction. For edge-preserving regularization, in addition to the smoothing parameter that balances data fidelity and regularization, one or more control parameters are typically incorporated to adjust the sensitivity of edge preservation by modifying the shape of the penalty function. Although there have been efforts to develop automated methods for tuning the hyperparameters in regularized PET reconstruction, the majority of these methods primarily focus on the smoothing parameter. However, it is challenging to obtain high-quality images without appropriately selecting the control parameters that adjust the edge preservation sensitivity. In this work, we propose a method to precisely tune the hyperparameters, which are initially set with a fixed value for the entire image, either manually or using an automated approach. Our core strategy involves adaptively adjusting the control parameter at each pixel, taking into account the degree of patch similarities calculated from the previous iteration within the pixel's neighborhood that is being updated. This approach allows our new method to integrate with a wide range of existing parameter-tuning techniques for edge-preserving regularization. Experimental results demonstrate that our proposed method effectively enhances the overall reconstruction accuracy across multiple image quality metrics, including peak signal-to-noise ratio, structural similarity, visual information fidelity, mean absolute error, root-mean-square error, and mean percentage error.

Adaptive diffusion priors for accelerated MRI reconstruction

Deep MRI reconstruction is commonly performed with conditional models that de-alias undersampled acquisitions to recover images consistent with fully-sampled data. Since conditional models are trained with knowledge of the imaging operator, they can show poor generalization across variable operators. Unconditional models instead learn generative image priors decoupled from the operator to improve reliability against domain shifts related to the imaging operator. Recent diffusion models are particularly promising given their high sample fidelity. Nevertheless, inference with a static image prior can perform suboptimally. Here we propose the first adaptive diffusion prior for MRI reconstruction, AdaDiff, to improve performance and reliability against domain shifts. AdaDiff leverages an efficient diffusion prior trained via adversarial mapping over large reverse diffusion steps. A two-phase reconstruction is executed following training: a rapid-diffusion phase that produces an initial reconstruction with the trained prior, and an adaptation phase that further refines the result by updating the prior to minimize data-consistency loss. Demonstrations on multi-contrast brain MRI clearly indicate that AdaDiff outperforms competing conditional and unconditional methods under domain shifts, and achieves superior or on par within-domain performance.

HFIST-Net: High-throughput fast iterative shrinkage thresholding network for accelerating MR image reconstruction

BACKGROUND AND OBJECTIVES

Compressed sensing (CS) is often used to accelerate magnetic resonance image (MRI) reconstruction from undersampled k-space data. A novelty deeply unfolded networks (DUNs) based method, designed by unfolding a traditional CS-MRI optimization algorithm into deep networks, can provide significantly faster reconstruction speeds than traditional CS-MRI methods while improving image quality.

METHODS

In this paper, we propose a High-Throughput Fast Iterative Shrinkage Thresholding Network (HFIST-Net) for reconstructing MR images from sparse measurements by combining traditional model-based CS techniques and data-driven deep learning methods. Specifically, the conventional Fast Iterative Shrinkage Thresholding Algorithm (FISTA) method is expanded as a deep network. To break the bottleneck of information transmission, a multi-channel fusion mechanism is proposed to improve the efficiency of information transmission between adjacent network stages. Moreover, a simple yet efficient channel attention block, called Gaussian context transformer (GCT), is proposed to improve the characterization capabilities of deep Convolutional Neural Network (CNN,) which utilizes Gaussian functions that satisfy preset relationships to achieve context feature excitation.

RESULTS

T1 and T2 brain MR images from the FastMRI dataset are used to validate the performance of the proposed HFIST-Net. The qualitative and quantitative results showed that our method is superior to those compared state-of-the-art unfolded deep learning networks.

CONCLUSIONS

The proposed HFIST-Net is capable of reconstructing more accurate MR image details from highly undersampled k-space data while maintaining fast computational speed.

ENSURE: A General Approach for Unsupervised Training of Deep Image Reconstruction Algorithms

Image reconstruction using deep learning algorithms offers improved reconstruction quality and lower reconstruction time than classical compressed sensing and model-based algorithms. Unfortunately, clean and fully sampled ground-truth data to train the deep networks is often unavailable in several applications, restricting the applicability of the above methods. We introduce a novel metric termed the ENsemble Stein's Unbiased Risk Estimate (ENSURE) framework, which can be used to train deep image reconstruction algorithms without fully sampled and noise-free images. The proposed framework is the generalization of the classical SURE and GSURE formulation to the setting where the images are sampled by different measurement operators, chosen randomly from a set. We evaluate the expectation of the GSURE loss functions over the sampling patterns to obtain the ENSURE loss function. We show that this loss is an unbiased estimate for the true mean-square error, which offers a better alternative to GSURE, which only offers an unbiased estimate for the projected error. Our experiments show that the networks trained with this loss function can offer reconstructions comparable to the supervised setting. While we demonstrate this framework in the context of MR image recovery, the ENSURE framework is generally applicable to arbitrary inverse problems.

Low-dose sinogram restoration enabled by conditional GAN with cross-domain regularization in SPECT imaging

In order to generate high-quality single-photon emission computed tomography (SPECT) images under low-dose acquisition mode, a sinogram denoising method was studied for suppressing random oscillation and enhancing contrast in the projection domain. A conditional generative adversarial network with cross-domain regularization (CGAN-CDR) is proposed for low-dose SPECT sinogram restoration. The generator stepwise extracts multiscale sinusoidal features from a low-dose sinogram, which are then rebuilt into a restored sinogram. Long skip connections are introduced into the generator, so that the low-level features can be better shared and reused, and the spatial and angular sinogram information can be better recovered. A patch discriminator is employed to capture detailed sinusoidal features within sinogram patches; thereby, detailed features in local receptive fields can be effectively characterized. Meanwhile, a cross-domain regularization is developed in both the projection and image domains. Projection-domain regularization directly constrains the generator via penalizing the difference between generated and label sinograms. Image-domain regularization imposes a similarity constraint on the reconstructed images, which can ameliorate the issue of ill-posedness and serves as an indirect constraint on the generator. By adversarial learning, the CGAN-CDR model can achieve high-quality sinogram restoration. Finally, the preconditioned alternating projection algorithm with total variation regularization is adopted for image reconstruction. Extensive numerical experiments show that the proposed model exhibits good performance in low-dose sinogram restoration. From visual analysis, CGAN-CDR performs well in terms of noise and artifact suppression, contrast enhancement and structure preservation, particularly in low-contrast regions. From quantitative analysis, CGAN-CDR has obtained superior results in both global and local image quality metrics. From robustness analysis, CGAN-CDR can better recover the detailed bone structure of the reconstructed image for a higher-noise sinogram. This work demonstrates the feasibility and effectiveness of CGAN-CDR in low-dose SPECT sinogram restoration. CGAN-CDR can yield significant quality improvement in both projection and image domains, which enables potential applications of the proposed method in real low-dose study.

Physics-/Model-Based and Data-Driven Methods for Low-Dose Computed Tomography: A survey

Since 2016, deep learning (DL) has advanced tomographic imaging with remarkable successes, especially in low-dose computed tomography (LDCT) imaging. Despite being driven by big data, the LDCT denoising and pure end-to-end reconstruction networks often suffer from the black box nature and major issues such as instabilities, which is a major barrier to apply deep learning methods in low-dose CT applications. An emerging trend is to integrate imaging physics and model into deep networks, enabling a hybridization of physics/model-based and data-driven elements. In this paper, we systematically review the physics/model-based data-driven methods for LDCT, summarize the loss functions and training strategies, evaluate the performance of different methods, and discuss relevant issues and future directions.

Linearized Analysis of Noise and Resolution for DL-Based Image Generation

Deep-learning (DL) based CT image generation methods are often evaluated using RMSE and SSIM. By contrast, conventional model-based image reconstruction (MBIR) methods are often evaluated using image properties such as resolution, noise, bias. Calculating such image properties requires time consuming Monte Carlo (MC) simulations. For MBIR, linearized analysis using first order Taylor expansion has been developed to characterize noise and resolution without MC simulations. This inspired us to investigate if linearization can be applied to DL networks to enable efficient characterization of resolution and noise. We used FBPConvNet as an example DL network and performed extensive numerical evaluations, including both computer simulations and real CT data. Our results showed that network linearization works well under normal exposure settings. For such applications, linearization can characterize image noise and resolutions without running MC simulations. We provide with this work the computational tools to implement network linearization. The efficiency and ease of implementation of network linearization can hopefully popularize the physics-related image quality measures for DL applications. Our methodology is general; it allows flexible compositions of DL nonlinear modules and linear operators such as filtered-backprojection (FBP). For the latter, we develop a generic method for computing the covariance images that is needed for network linearization.

Deep image and feature prior algorithm based on U-ConformerNet structure

PURPOSE

The reconstruction performance of the deep image prior (DIP) approach is limited by the conventional convolutional layer structure and it is difficult to enhance its potential. In order to improve the quality of image reconstruction and suppress artifacts, we propose a DIP algorithm with better performance, and verify its superiority in the latest case.

METHODS

We construct a new U-ConformerNet structure as the DIP algorithm's network, replacing the traditional convolutional layer-based U-net structure, and introduce the 'lpips' deep network based feature distance regularization method. Our algorithm can switch between supervised and unsupervised modes at will to meet different needs.

RESULTS

The reconstruction was performed on the low dose CT dataset (LoDoPaB). Our algorithm attained a PSNR of more than 35 dB under unsupervised conditions, and the PSNR under the supervised condition is greater than 36 dB. Both of which are better than the performance of the DIP-TV. Furthermore, the accuracy of this method is positively connected with the quality of the a priori image with the help of deep networks. In terms of noise eradication and artifact suppression, the DIP algorithm with U-ConformerNet structure outperforms the standard DIP method based on convolutional structure.

CONCLUSIONS

It is known by experimental verification that, in unsupervised mode, the algorithm improves the output PSNR by at least 2-3 dB when compared to the DIP-TV algorithm (proposed in 2020). In supervised mode, our algorithm approaches that of the state-of-the-art end-to-end deep learning algorithms.

TextureWGAN: texture preserving WGAN with multitask regularizer for computed tomography inverse problems

Purpose

This paper presents a deep learning (DL) based method called TextureWGAN. It is designed to preserve image texture while maintaining high pixel fidelity for computed tomography (CT) inverse problems. Over-smoothed images by postprocessing algorithms have been a well-known problem in the medical imaging industry. Therefore, our method tries to solve the over-smoothing problem without compromising pixel fidelity.

Approach

The TextureWGAN extends from Wasserstein GAN (WGAN). The WGAN can create an image that looks like a genuine image. This aspect of the WGAN helps preserve image texture. However, an output image from the WGAN is not correlated to the corresponding ground truth image. To solve this problem, we introduce the multitask regularizer (MTR) to the WGAN framework to make a generated image highly correlated to the corresponding ground truth image so that the TextureWGAN can achieve high-level pixel fidelity. The MTR is capable of using multiple objective functions. In this research, we adopt a mean squared error (MSE) loss to maintain pixel fidelity. We also use a perception loss to improve the look and feel of result images. Furthermore, the regularization parameters in the MTR are trained along with generator network weights to maximize the performance of the TextureWGAN generator.

Results

The proposed method was evaluated in CT image reconstruction applications in addition to super-resolution and image-denoising applications. We conducted extensive qualitative and quantitative evaluations. We used PSNR and SSIM for pixel fidelity analysis and the first-order and the second-order statistical texture analysis for image texture. The results show that the TextureWGAN is more effective in preserving image texture compared with other well-known methods such as the conventional CNN and nonlocal mean filter (NLM). In addition, we demonstrate that TextureWGAN can achieve competitive pixel fidelity performance compared with CNN and NLM. The CNN with MSE loss can attain high-level pixel fidelity, but it often damages image texture.

Conclusions

TextureWGAN can preserve image texture while maintaining pixel fidelity. The MTR is not only helpful to stabilize the TextureWGAN's generator training but also maximizes the generator performance.

Deep Gauss-Newton for phase retrieval

We propose the deep Gauss-Newton (DGN) algorithm. The DGN allows one to take into account the knowledge of the forward model in a deep neural network by unrolling a Gauss-Newton optimization method. No regularization or step size needs to be chosen; they are learned through convolutional neural networks. The proposed algorithm does not require an initial reconstruction and is able to retrieve simultaneously the phase and absorption from a single-distance diffraction pattern. The DGN method was applied to both simulated and experimental data and permitted large improvements of the reconstruction error and of the resolution compared with a state-of-the-art iterative method and another neural-network-based reconstruction algorithm.

A total variation prior unrolling approach for computed tomography reconstruction

BACKGROUND

With the rapid development of deep learning technology, deep neural networks can effectively enhance the performance of computed tomography (CT) reconstructions. One kind of commonly used method to construct CT reconstruction networks is to unroll the conventional iterative reconstruction (IR) methods to convolutional neural networks (CNNs). However, most unrolling methods primarily unroll the fidelity term of IR methods to CNNs, without unrolling the prior terms. The prior terms are always directly replaced by neural networks.

PURPOSE

In conventional IR methods, the prior terms play a vital role in improving the visual quality of reconstructed images. Unrolling the hand-crafted prior terms to CNNs may provide a more specialized unrolling approach to further improve the performance of CT reconstruction. In this work, a primal-dual network (PD-Net) was proposed by unrolling both the data fidelity term and the total variation (TV) prior term, which effectively preserves the image edges and textures in the reconstructed images.

METHODS

By further deriving the Chambolle-Pock (CP) algorithm instance for CT reconstruction, we discovered that the TV prior updates the reconstructed images with its divergences in each iteration of the solution process. Based on this discovery, CNNs were applied to yield the divergences of the feature maps for the reconstructed image generated in each iteration. Additionally, a loss function was applied to the predicted divergences of the reconstructed image to guarantee that the CNNs' results were the divergences of the corresponding feature maps in the iteration. In this manner, the proposed CNNs seem to play the same roles in the PD-Net as the TV prior in the IR methods. Thus, the TV prior in the CP algorithm instance can be directly unrolled to CNNs.

RESULTS

The datasets from the Low-Dose CT Image and Projection Data and the Piglet dataset were employed to assess the effectiveness of our proposed PD-Net. Compared with conventional CT reconstruction methods, our proposed method effectively preserves the structural and textural information in reference to ground truth.

CONCLUSIONS

The experimental results show that our proposed PD-Net framework is feasible for the implementation of CT reconstruction tasks. Owing to the promising results yielded by our proposed neural network, this study is intended to inspire further development of unrolling approaches by enabling the direct unrolling of hand-crafted prior terms to CNNs.

Fast iterative reconstruction for photoacoustic tomography using learned physical model: Theoretical validation

Iterative reconstruction has demonstrated superior performance in medical imaging under compressed, sparse, and limited-view sensing scenarios. However, iterative reconstruction algorithms are slow to converge and rely heavily on hand-crafted parameters to achieve good performance. Many iterations are usually required to reconstruct a high-quality image, which is computationally expensive due to repeated evaluations of the physical model. While learned iterative reconstruction approaches such as model-based learning (MBLr) can reduce the number of iterations through convolutional neural networks, it still requires repeated evaluations of the physical models at each iteration. Therefore, the goal of this study is to develop a Fast Iterative Reconstruction (FIRe) algorithm that incorporates a learned physical model into the learned iterative reconstruction scheme to further reduce the reconstruction time while maintaining robust reconstruction performance. We also propose an efficient training scheme for FIRe, which releases the enormous memory footprint required by learned iterative reconstruction methods through the concept of recursive training. The results of our proposed method demonstrate comparable reconstruction performance to learned iterative reconstruction methods with a 9x reduction in computation time and a 620x reduction in computation time compared to variational reconstruction.

RISING: A new framework for model-based few-view CT image reconstruction with deep learning

Medical image reconstruction from low-dose tomographic data is an active research field, recently revolutionized by the advent of deep learning. In fact, deep learning typically yields superior results than classical optimization approaches, but unstable results have been reported both numerically and theoretically in the literature. This paper proposes RISING, a new framework for sparse-view tomographic image reconstruction combining an early-stopped Rapid Iterative Solver with a subsequent Iteration Network-based Gaining step. In our two-step approach, the first phase executes very few iterations of a regularized model-based algorithm, whereas the second step completes the missing iterations by means of a convolutional neural network. The proposed method is ground-truth free; it exploits the computational speed and flexibility of a data-driven approach, but it also imposes sparsity constraints to the solution as in the model-based setting. Experiments performed both on a digitally created and on a real abdomen data set confirm the numerical and visual accuracy of the reconstructed RISING images in short computational times. These features make the framework promising to be used on real systems for clinical exams.

Solving Inverse Problems With Deep Neural Networks - Robustness Included?

In the past five years, deep learning methods have become state-of-the-art in solving various inverse problems. Before such approaches can find application in safety-critical fields, a verification of their reliability appears mandatory. Recent works have pointed out instabilities of deep neural networks for several image reconstruction tasks. In analogy to adversarial attacks in classification, it was shown that slight distortions in the input domain may cause severe artifacts. The present article sheds new light on this concern, by conducting an extensive study of the robustness of deep-learning-based algorithms for solving underdetermined inverse problems. This covers compressed sensing with Gaussian measurements as well as image recovery from Fourier and Radon measurements, including a real-world scenario for magnetic resonance imaging (using the NYU-fastMRI dataset). Our main focus is on computing adversarial perturbations of the measurements that maximize the reconstruction error. A distinctive feature of our approach is the quantitative and qualitative comparison with total-variation minimization, which serves as a provably robust reference method. In contrast to previous findings, our results reveal that standard end-to-end network architectures are not only resilient against statistical noise, but also against adversarial perturbations. All considered networks are trained by common deep learning techniques, without sophisticated defense strategies.

Motion grading of high-resolution quantitative computed tomography supported by deep convolutional neural networks

Image quality degradation due to subject motion confounds the precision and reproducibility of measurements of bone density, morphology and mechanical properties from high-resolution peripheral quantitative computed tomography (HR-pQCT). Time-consuming operator-based scoring of motion artefacts remains the gold standard to determine the degree of acceptable motion. However, due to the subjectiveness of manual grading, HR-pQCT scans of poor quality, which cannot be used for analysis, may be accepted upon initial review, leaving patients with incomplete or inaccurate imaging results. Convolutional Neural Networks (CNNs) enable fast image analysis with relatively few pre-processing requirements in an operator-independent and fully automated way for image classification tasks. This study aimed to develop a CNN that can predict motion scores from HR-pQCT images, while also being aware of uncertain predictions that require manual confirmation. The CNN calculated motion scores within seconds and achieved a high F1-score (86.8 ± 2.8 %), with good precision (87.5 ± 2.7 %), recall (86.7 ± 2.9 %) and a substantial agreement with the ground truth measured by Cohen's kappa (κ = 68.6 ± 6.2 %); motion scores of the test dataset were predicted by the algorithm with comparable accuracy, precision, sensitivity and agreement as by the operators (p > 0.05). This post-processing approach may be used to assess the effect of motion scores on microstructural analysis and can be immediately implemented into clinical protocols, significantly reducing the time for quality assessment and control of HR-pQCT scans.

Unrolled primal-dual networks for lensless cameras

Conventional models for lensless imaging assume that each measurement results from convolving a given scene with a single experimentally measured point-spread function. These models fail to simulate lensless cameras truthfully, as these models do not account for optical aberrations or scenes with depth variations. Our work shows that learning a supervised primal-dual reconstruction method results in image quality matching state of the art in the literature without demanding a large network capacity. We show that embedding learnable forward and adjoint models improves the reconstruction quality of lensless images (+5dB PSNR) compared to works that assume a fixed point-spread function.

Superiorization-inspired unrolled SART algorithm with U-Net generated perturbations for sparse-view and limited-angle CT reconstruction

Objective.Unrolled algorithms are a promising approach for reconstruction of CT images in challenging scenarios, such as low-dose, sparse-view and limited-angle imaging. In an unrolled algorithm, a fixed number of iterations of a reconstruction method are unrolled into multiple layers of a neural network, and interspersed with trainable layers. The entire network is then trained end-to-end in a supervised fashion, to learn an appropriate regularizer from training data. In this paper we propose a novel unrolled algorithm, and compare its performance with several other approaches on sparse-view and limited-angle CT.Approach.The proposed algorithm is inspired by the superiorization methodology, an optimization heuristic in which iterates of a feasibility-seeking method are perturbed between iterations, typically using descent directions of a model-based penalty function. Our algorithm instead uses a modified U-net architecture to introduce the perturbations, allowing a network to learn beneficial perturbations to the image at various stages of the reconstruction, based on the training data.Main Results.In several numerical experiments modeling sparse-view and limited angle CT scenarios, the algorithm provides excellent results. In particular, it outperforms several competing unrolled methods in limited-angle scenarios, while providing comparable or better performance on sparse-view scenarios.Significance.This work represents a first step towards exploiting the power of deep learning within the superiorization methodology. Additionally, it studies the effect of network architecture on the performance of unrolled methods, as well as the effectiveness of the unrolled approach on both limited-angle CT, where previous studies have primarily focused on the sparse-view and low-dose cases.

Noise Characteristics Modeled Unsupervised Network for Robust CT Image Reconstruction

Deep learning (DL)-based methods show great potential in computed tomography (CT) imaging field. The DL-based reconstruction methods are usually evaluated on the training and testing datasets which are obtained from the same distribution, i.e., the same CT scan protocol (i.e., the region setting, kVp, mAs, etc.). In this work, we focus on analyzing the robustness of the DL-based methods against protocol-specific distribution shifts (i.e., the training and testing datasets are from different region settings, different kVp settings, or different mAs settings, respectively). The results show that the DL-based reconstruction methods are sensitive to the protocol-specific perturbations which can be attributed to the noise distribution shift between the training and testing datasets. Based on these findings, we presented a low-dose CT reconstruction method using an unsupervised strategy with the consideration of noise distribution to address the issue of protocol-specific perturbations. Specifically, unpaired sinogram data is enrolled into the network training, which represents unique information for specific imaging protocol, and a Gaussian mixture model (GMM) is introduced to characterize the noise distribution in CT images. It can be termed as GMM based unsupervised CT reconstruction network (GMM-unNet) method. Moreover, an expectation-maximization algorithm is designed to optimize the presented GMM-unNet method. Extensive experiments are performed on three datasets from different scan protocols, which demonstrate that the presented GMM-unNet method outperforms the competing methods both qualitatively and quantitatively.

TransCS: A Transformer-Based Hybrid Architecture for Image Compressed Sensing

Well-known compressed sensing (CS) is widely used in image acquisition and reconstruction. However, accurately reconstructing images from measurements at low sampling rates remains a considerable challenge. In this paper, we propose a novel Transformer-based hybrid architecture (dubbed TransCS) to achieve high-quality image CS. In the sampling module, TransCS adopts a trainable sensing matrix strategy that gains better image reconstruction by learning the structural information from the training images. In the reconstruction module, inspired by the powerful long-distance dependence modelling capacity of the Transformer, a customized iterative shrinkage-thresholding algorithm (ISTA)-based Transformer backbone that iteratively works with gradient descent and soft threshold operation is designed to model the global dependency among image subblocks. Moreover, the auxiliary convolutional neural network (CNN) is introduced to capture the local features of images. Therefore, the proposed hybrid architecture that integrates the customized ISTA-based Transformer backbone with CNN can gain high-performance reconstruction for image compressed sensing. The experimental results demonstrate that our proposed TransCS obtains superior reconstruction quality and noise robustness on several public benchmark datasets compared with other state-of-the-art methods. Our code is available on TransCS.

Deep reinforcement learning and its applications in medical imaging and radiation therapy: a survey

Reinforcement learning takes sequential decision-making approaches by learning the policy through trial and error based on interaction with the environment. Combining deep learning and reinforcement learning can empower the agent to learn the interactions and the distribution of rewards from state-action pairs to achieve effective and efficient solutions in more complex and dynamic environments. Deep reinforcement learning (DRL) has demonstrated astonishing performance in surpassing the human-level performance in the game domain and many other simulated environments. This paper introduces the basics of reinforcement learning and reviews various categories of DRL algorithms and DRL models developed for medical image analysis and radiation treatment planning optimization. We will also discuss the current challenges of DRL and approaches proposed to make DRL more generalizable and robust in a real-world environment. DRL algorithms, by fostering the designs of the reward function, agents interactions and environment models, can resolve the challenges from scarce and heterogeneous annotated medical image data, which has been a major obstacle to implementing deep learning models in the clinic. DRL is an active research area with enormous potential to improve deep learning applications in medical imaging and radiation therapy planning.

ERA-WGAT: Edge-enhanced residual autoencoder with a window-based graph attention convolutional network for low-dose CT denoising

Computed tomography (CT) has become a powerful tool for medical diagnosis. However, minimizing X-ray radiation risk for the patient poses significant challenges to obtain suitable low dose CT images. Although various low-dose CT methods using deep learning techniques have produced impressive results, convolutional neural network based methods focus more on local information and hence are very limited for non-local information extraction. This paper proposes ERA-WGAT, a residual autoencoder incorporating an edge enhancement module that performs convolution with eight types of learnable operators providing rich edge information and a window-based graph attention convolutional network that combines static and dynamic attention modules to explore non-local self-similarity. We use the compound loss function that combines MSE loss and multi-scale perceptual loss to mitigate the over-smoothing problem. Compared with current low-dose CT denoising methods, ERA-WGAT confirmed superior noise suppression and perceived image quality.

A dual-domain neural network based on sinogram synthesis for sparse-view CT reconstruction

OBJECTIVE

The dual-domain deep learning-based reconstruction techniques have enjoyed many successful applications in the field of medical image reconstruction. Applying the analytical reconstruction based operator to transfer the data from the projection domain to the image domain, the dual-domain techniques may suffer from the insufficient suppression or removal of streak artifacts in areas with the missing view data, when addressing the sparse-view reconstruction problems. In this work, to overcome this problem, an intelligent sinogram synthesis based back-projection network (iSSBP-Net) was proposed for sparse-view computed tomography (CT) reconstruction. In the iSSBP-Net method, a convolutional neural network (CNN) was involved in the dual-domain method to inpaint the missing view data in the sinogram before CT reconstruction.

METHODS

The proposed iSSBP-Net method fused a sinogram synthesis sub-network (SS-Net), a sinogram filter sub-network (SF-Net), a back-projection layer, and a post-CNN into an end-to-end network. Firstly, to inpaint the missing view data, the SS-Net employed a CNN to synthesize the full-view sinogram in the projection domain. Secondly, to improve the visual quality of the sparse-view CT images, the synthesized sinogram was filtered by a CNN. Thirdly, the filtered sinogram was brought into the image domain through the back-projection layer. Finally, to yield images of high visual sensitivity, the post-CNN was applied to restore the desired images from the outputs of the back-projection layer.

RESULTS

The numerical experiments demonstrate that the proposed iSSBP-Net is superior to all competing algorithms under different scanning condintions for sparse-view CT reconstruction. Compared to the competing algorithms, the proposed iSSBP-Net method improved the peak signal-to-noise ratio of the reconstructed images about 1.21 dB, 0.26 dB, 0.01 dB, and 0.37 dB under the scanning conditions of 360, 180, 90, and 60 views, respectively.

CONCLUSION

The promising reconstruction results indicate that involving the SS-Net in the dual-domain method is could be an effective manner to suppress or remove the streak artifacts in sparse-view CT images. Due to the promising results reconstructed by the iSSBP-Net method, this study is intended to inspire the further development of sparse-view CT reconstruction by involving a SS-Net in the dual-domain method.

Iterative tomographic reconstruction with TV prior for low-dose CBCT dental imaging

Objective.Cone-beam computed tomography is becoming more and more popular in applications such as 3D dental imaging. Iterative methods compared to the standard Feldkamp algorithm have shown improvements in image quality of reconstruction of low-dose acquired data despite their long computing time. An interesting aspect of iterative methods is their ability to include prior information such as sparsity-constraint. While a large panel of optimization algorithms along with their adaptation to tomographic problems are available, they are mainly studied on 2D parallel or fan-beam data. The issues raised by 3D CBCT and moreover by truncated projections are still poorly understood.Approach.We compare different carefully designed optimization schemes in the context of realistic 3D dental imaging. Besides some known algorithms, SIRT-TV and MLEM, we investigate the primal-dual hybrid gradient (PDHG) approach and a newly proposed MLEM-TV optimizer. The last one is alternating EM steps and TV-denoising, combination not yet investigated for CBCT. Experiments are performed on both simulated data from a 3D jaw phantom and data acquired with a dental clinical scanner.Main results.With some adaptations to the specificities of CBCT operators, PDHG and MLEM-TV algorithms provide the best reconstruction quality. These results were obtained by comparing the full-dose image with a low-dose image and an ultra low-dose image.Significance.The convergence speed of the original iterative methods is hampered by the conical geometry and significantly reduced compared to parallel geometries. We promote the pre-conditioned version of PDHG and we propose a pre-conditioned version of the MLEM-TV algorithm. To the best of our knowledge, this is the first time PDHG and convergent MLEM-TV algorithms are evaluated on experimental dental CBCT data, where constraints such as projection truncation and presence of metal have to be jointly overcome.

Unsupervised knowledge-transfer for learned image reconstruction

Deep learning-based image reconstruction approaches have demonstrated impressive empirical performance in many imaging modalities. These approaches usually require a large amount of high-quality paired training data, which is often not available in medical imaging. To circumvent this issue we develop a novel unsupervised knowledge-transfer paradigm for learned reconstruction within a Bayesian framework. The proposed approach learns a reconstruction network in two phases. The first phase trains a reconstruction network with a set of ordered pairs comprising of ground truth images of ellipses and the corresponding simulated measurement data. The second phase fine-tunes the pretrained network to more realistic measurement data without supervision. By construction, the framework is capable of delivering predictive uncertainty information over the reconstructed image. We present extensive experimental results on low-dose and sparse-view computed tomography showing that the approach is competitive with several state-of-the-art supervised and unsupervised reconstruction techniques. Moreover, for test data distributed differently from the training data, the proposed framework can significantly improve reconstruction quality not only visually, but also quantitatively in terms of PSNR and SSIM, when compared with learned methods trained on the synthetic dataset only.

A cascade of preconditioned conjugate gradient networks for accelerated magnetic resonance imaging

BACKGROUND AND OBJECTIVE

Recent unfolding based compressed sensing magnetic resonance imaging (CS-MRI) methods only reinterpret conventional CS-MRI optimization algorithms and, consequently, inherit the weaknesses of the alternating optimization strategy. In order to avoid the structural complexity of the alternating optimization strategy and achieve better reconstruction performance, we propose to directly optimize the ℓ₁ regularized convex optimization problem using a deep learning approach.

METHOD

In order to achieve direct optimization, a system of equations solving the ℓ₁ regularized optimization problem is constructed from the optimality conditions of a novel primal-dual form proposed for the effective training of the sparsifying transform. The optimal solution is obtained by a cascade of unfolding networks of the preconditioned conjugate gradient (PCG) algorithm trained to minimize the mean element-wise absolute difference (ℓ₁ loss) between the terminal output and ground truth image in an end-to-end manner. The performance of the proposed method was compared with that of U-Net, PD-Net, ISTA-Net+, and the recently proposed projection-based cascaded U-Net, using single-coil knee MR images of the fastMRI dataset.

RESULTS

In our experiment, the proposed network outperformed existing unfolding-based networks and the complex version of U-Net in several subsampling scenarios. In particular, when using the random Cartesian subsampling mask with 25 % sampling rate, the proposed model outperformed PD-Net by 0.76 dB, ISTA-Net+ by 0.43 dB, and U-Net by 1.21 dB on the positron density without suppression (PD) dataset in term of peak signal to noise ratio. In comparison with the projection-based cascade U-Net, the proposed algorithm achieved approximately the same performance when the sampling rate was 25% with only 1.62% number of network parameters at the cost of a longer reconstruction time (approximately twice).

CONCLUSION

A cascade of unfolding networks of the PCG algorithm was proposed to directly optimize the ℓ₁ regularized CS-MRI optimization problem. The proposed network achieved improved reconstruction performance compared with U-Net, PD-Net, and ISTA-Net+, and achieved approximately the same performance as the projection-based cascaded U-Net while using significantly fewer network parameters.

Self-Training Strategy Based on Finite Element Method for Adaptive Bioluminescence Tomography Reconstruction

Bioluminescence tomography (BLT) is a promising pre-clinical imaging technique for a wide variety of biomedical applications, which can non-invasively reveal functional activities inside living animal bodies through the detection of visible or near-infrared light produced by bioluminescent reactions. Recently, reconstruction approaches based on deep learning have shown great potential in optical tomography modalities. However, these reports only generate data with stationary patterns of constant target number, shape, and size. The neural networks trained by these data sets are difficult to reconstruct the patterns outside the data sets. This will tremendously restrict the applications of deep learning in optical tomography reconstruction. To address this problem, a self-training strategy is proposed for BLT reconstruction in this paper. The proposed strategy can fast generate large-scale BLT data sets with random target numbers, shapes, and sizes through an algorithm named random seed growth algorithm and the neural network is automatically self-trained. In addition, the proposed strategy uses the neural network to build a map between photon densities on surface and inside the imaged object rather than an end-to-end neural network that directly infers the distribution of sources from the photon density on surface. The map of photon density is further converted into the distribution of sources through the multiplication with stiffness matrix. Simulation, phantom, and mouse studies are carried out. Results show the availability of the proposed self-training strategy.

GAN-TL: Generative Adversarial Networks with Transfer Learning for MRI Reconstruction

Generative adversarial networks (GAN), which are fueled by deep learning, are an efficient technique for image reconstruction using under-sampled MR data. In most cases, the performance of a particular model’s reconstruction must be improved by using a substantial proportion of the training data. However, gathering tens of thousands of raw patient data for training the model in actual clinical applications is difficult because retaining k-space data is not customary in the clinical process. Therefore, it is imperative to increase the generalizability of a network that was created using a small number of samples as quickly as possible. This research explored two unique applications based on deep learning-based GAN and transfer learning. Seeing as MRI reconstruction procedures go for brain and knee imaging, the proposed method outperforms current techniques in terms of signal-to-noise ratio (PSNR) and structural similarity index (SSIM). As compared to the results of transfer learning for the brain and knee, using a smaller number of training cases produced superior results, with acceleration factor (AF) 2 (for brain PSNR (39.33); SSIM (0.97), for knee PSNR (35.48); SSIM (0.90)) and AF 4 (for brain PSNR (38.13); SSIM (0.95), for knee PSNR (33.95); SSIM (0.86)). The approach that has been described would make it easier to apply future models for MRI reconstruction without necessitating the acquisition of vast imaging datasets.

A review on the application of deep learning for CT reconstruction, bone segmentation and surgical planning in oral and maxillofacial surgery

Computer-assisted surgery (CAS) allows clinicians to personalize treatments and surgical interventions and has therefore become an increasingly popular treatment modality in maxillofacial surgery. The current maxillofacial CAS consists of three main steps: (1) CT image reconstruction, (2) bone segmentation, and (3) surgical planning. However, each of these three steps can introduce errors that can heavily affect the treatment outcome. As a consequence, tedious and time-consuming manual post-processing is often necessary to ensure that each step is performed adequately. One way to overcome this issue is by developing and implementing neural networks (NNs) within the maxillofacial CAS workflow. These learning algorithms can be trained to perform specific tasks without the need for explicitly defined rules. In recent years, an extremely large number of novel NN approaches have been proposed for a wide variety of applications, which makes it a difficult task to keep up with all relevant developments. This study therefore aimed to summarize and review all relevant NN approaches applied for CT image reconstruction, bone segmentation, and surgical planning. After full text screening, 76 publications were identified: 32 focusing on CT image reconstruction, 33 focusing on bone segmentation and 11 focusing on surgical planning. Generally, convolutional NNs were most widely used in the identified studies, although the multilayer perceptron was most commonly applied in surgical planning tasks. Moreover, the drawbacks of current approaches and promising research avenues are discussed.

Iterative phase contrast CT reconstruction with novel tomographic operator and data-driven prior

Breast cancer remains the most prevalent malignancy in women in many countries around the world, thus calling for better imaging technologies to improve screening and diagnosis. Grating interferometry (GI)-based phase contrast X-ray CT is a promising technique which could make the transition to clinical practice and improve breast cancer diagnosis by combining the high three-dimensional resolution of conventional CT with higher soft-tissue contrast. Unfortunately though, obtaining high-quality images is challenging. Grating fabrication defects and photon starvation lead to high noise amplitudes in the measured data. Moreover, the highly ill-conditioned differential nature of the GI-CT forward operator renders the inversion from corrupted data even more cumbersome. In this paper, we propose a novel regularized iterative reconstruction algorithm with an improved tomographic operator and a powerful data-driven regularizer to tackle this challenging inverse problem. Our algorithm combines the L-BFGS optimization scheme with a data-driven prior parameterized by a deep neural network. Importantly, we propose a novel regularization strategy to ensure that the trained network is non-expansive, which is critical for the convergence and stability analysis we provide. We empirically show that the proposed method achieves high quality images, both on simulated data as well as on real measurements.

LRR-CED: low-resolution reconstruction-aware convolutional encoder–decoder network for direct sparse-view CT image reconstruction

Objective. Sparse-view computed tomography (CT) reconstruction has been at the forefront of research in medical imaging. Reducing the total x-ray radiation dose to the patient while preserving the reconstruction accuracy is a big challenge. The sparse-view approach is based on reducing the number of rotation angles, which leads to poor quality reconstructed images as it introduces several artifacts. These artifacts are more clearly visible in traditional reconstruction methods like the filtered-backprojection (FBP) algorithm. Approach. Over the years, several model-based iterative and more recently deep learning-based methods have been proposed to improve sparse-view CT reconstruction. Many deep learning-based methods improve FBP-reconstructed images as a post-processing step. In this work, we propose a direct deep learning-based reconstruction that exploits the information from low-dimensional scout images, to learn the projection-to-image mapping. This is done by concatenating FBP scout images at multiple resolutions in the decoder part of a convolutional encoder–decoder (CED). Main results. This approach is investigated on two different networks, based on Dense Blocks and U-Net to show that a direct mapping can be learned from a sinogram to an image. The results are compared to two post-processing deep learning methods (FBP-ConvNet and DD-Net) and an iterative method that uses a total variation (TV) regularization. Significance. This work presents a novel method that uses information from both sinogram and low-resolution scout images for sparse-view CT image reconstruction. We also generalize this idea by demonstrating results with two different neural networks. This work is in the direction of exploring deep learning across the various stages of the image reconstruction pipeline involving data correction, domain transfer and image improvement.

The Complex-valued PD-net for MRI reconstruction of knee images

MRI reconstruction is the fundamental task of obtaining diagnostic quality images from MRI sensor data and is an active area of research for improving accuracy, speed and memory requirements of the process. Complex-valued neural networks have previously achieved superior MRI reconstructions compared to real-valued nets. But those works operated in the image domain to denoise poor quality reconstructions of the raw sensor (k-space) data. Also small-scale or proprietary datasets with few clinical images or raw k-space volumes were used in these works, and none of the works use publicly available large-scale raw k-space datasets. Recent studies have shown that cross-domain neural networks for MRI reconstruction, or networks which leverage information from both k-space and image domains, have better potential than single-domain networks which operate only in one domain. We study the effects of complex-valued operations on a top-performing cross-domain neural network for MRI reconstruction called the Primal-Dual net, or PD-net. The PD-net is a fully convolutional architecture that takes input as raw k-space data and outputs the reconstructions, thus performing both the inversion and denoising tasks. We experiment with the publicly available, large-scale fastMRI single-coil knee dataset having 973 train volumes and 199 validation volumes. Our proposed method (Complex PD-net) achieves PSNR and SSIM of 33.3 dB and 0.8033 respectively, compared to 32.13 dB and 0.728 obtained by PD-net. Our Complex PD-net achieves 10.3% higher SSIM with just over 50% of the total parameters w.r.t. the SOTA methodology.

Unsupervised MRI Reconstruction via Zero-Shot Learned Adversarial Transformers

Supervised reconstruction models are characteristically trained on matched pairs of undersampled and fully-sampled data to capture an MRI prior, along with supervision regarding the imaging operator to enforce data consistency. To reduce supervision requirements, the recent deep image prior framework instead conjoins untrained MRI priors with the imaging operator during inference. Yet, canonical convolutional architectures are suboptimal in capturing long-range relationships, and priors based on randomly initialized networks may yield suboptimal performance. To address these limitations, here we introduce a novel unsupervised MRI reconstruction method based on zero-Shot Learned Adversarial TransformERs (SLATER). SLATER embodies a deep adversarial network with cross-attention transformers to map noise and latent variables onto coil-combined MR images. During pre-training, this unconditional network learns a high-quality MRI prior in an unsupervised generative modeling task. During inference, a zero-shot reconstruction is then performed by incorporating the imaging operator and optimizing the prior to maximize consistency to undersampled data. Comprehensive experiments on brain MRI datasets clearly demonstrate the superior performance of SLATER against state-of-the-art unsupervised methods.

NC-PDNet: A Density-Compensated Unrolled Network for 2D and 3D Non-Cartesian MRI Reconstruction

Deep Learning has become a very promising avenue for magnetic resonance image (MRI) reconstruction. In this work, we explore the potential of unrolled networks for non-Cartesian acquisition settings. We design the NC-PDNet (Non-Cartesian Primal Dual Netwok), the first density-compensated (DCp) unrolled neural network, and validate the need for its key components via an ablation study. Moreover, we conduct some generalizability experiments to test this network in out-of-distribution settings, for example training on knee data and validating on brain data. The results show that NC-PDNet outperforms baseline (U-Net, Deep image prior) models both visually and quantitatively in all settings. In particular, in the 2D multi-coil acquisition scenario, the NC-PDNet provides up to a 1.2 dB improvement in peak signal-to-noise ratio (PSNR) over baseline networks, while also allowing a gain of at least 1dB in PSNR in generalization settings. We provide the open-source implementation of NC-PDNet, and in particular the Non-uniform Fourier Transform in TensorFlow, tested on 2D multi-coil and 3D single-coil k-space data.

Assessment of data consistency through cascades of independently recurrent inference machines for fast and robust accelerated MRI reconstruction

Objective. Machine Learning methods can learn how to reconstruct magnetic resonance images (MRI) and thereby accelerate acquisition, which is of paramount importance to the clinical workflow. Physics-informed networks incorporate the forward model of accelerated MRI reconstruction in the learning process. With increasing network complexity, robustness is not ensured when reconstructing data unseen during training. We aim to embed data consistency (DC) in deep networks while balancing the degree of network complexity. While doing so, we will assess whether either explicit or implicit enforcement of DC in varying network architectures is preferred to optimize performance. Approach. We propose a scheme called Cascades of Independently Recurrent Inference Machines (CIRIM) to assess DC through unrolled optimization. Herein we assess DC both implicitly by gradient descent and explicitly by a designed term. Extensive comparison of the CIRIM to compressed sensing as well as other Machine Learning methods is performed: the End-to-End Variational Network (E2EVN), CascadeNet, KIKINet, LPDNet, RIM, IRIM, and UNet. Models were trained and evaluated on T1-weighted and FLAIR contrast brain data, and T2-weighted knee data. Both 1D and 2D undersampling patterns were evaluated. Robustness was tested by reconstructing 7.5× prospectively undersampled 3D FLAIR MRI data of multiple sclerosis (MS) patients with white matter lesions. Main results. The CIRIM performed best when implicitly enforcing DC, while the E2EVN required an explicit DC formulation. Through its cascades, the CIRIM was able to score higher on structural similarity and PSNR compared to other methods, in particular under heterogeneous imaging conditions. In reconstructing MS patient data, prospectively acquired with a sampling pattern unseen during model training, the CIRIM maintained lesion contrast while efficiently denoising the images. Significance. The CIRIM showed highly promising generalization capabilities maintaining a very fair trade-off between reconstructed image quality and fast reconstruction times, which is crucial in the clinical workflow.

The use of deep learning methods in low-dose computed tomography image reconstruction: a systematic review

Conventional reconstruction techniques, such as filtered back projection (FBP) and iterative reconstruction (IR), which have been utilised widely in the image reconstruction process of computed tomography (CT) are not suitable in the case of low-dose CT applications, because of the unsatisfying quality of the reconstructed image and inefficient reconstruction time. Therefore, as the demand for CT radiation dose reduction continues to increase, the use of artificial intelligence (AI) in image reconstruction has become a trend that attracts more and more attention. This systematic review examined various deep learning methods to determine their characteristics, availability, intended use and expected outputs concerning low-dose CT image reconstruction. Utilising the methodology of Kitchenham and Charter, we performed a systematic search of the literature from 2016 to 2021 in Springer, Science Direct, arXiv, PubMed, ACM, IEEE, and Scopus. This review showed that algorithms using deep learning technology are superior to traditional IR methods in noise suppression, artifact reduction and structure preservation, in terms of improving the image quality of low-dose reconstructed images. In conclusion, we provided an overview of the use of deep learning approaches in low-dose CT image reconstruction together with their benefits, limitations, and opportunities for improvement.

Cycle-consistent learning-based hybrid iterative reconstruction for whole-body PET imaging

Objective. To develop a cycle-consistent learning-based hybrid iterative reconstruction (IR) method that takes only slightly longer than analytic reconstruction, while pursuing the image resolution and tumor quantification achievable by IR for whole-body PET imaging. Approach. We backproject the raw positron emission tomography (PET) data to generate a blurred activity distribution. From the backprojection to the IR label, a reconstruction mapping that approximates the deblurring filters for the point spread function and the physical effects of the PET system is unrolled to a neural network with stacked convolutional layers. By minimizing the cycle-consistent loss, we train the reconstruction and inverse mappings simultaneously. Main results. In phantom study, the proposed method results in an absolute relative error (RE) of the mean activity of 4.0% ± 0.7% in the largest hot sphere, similar to the RE of the full-count IR and significantly smaller than that obtained by CycleGAN postprocessing. Achieving a noise reduction of 48.1% ± 0.5% relative to the low-count IR, the proposed method demonstrates advantages over the low-count IR and CycleGAN in terms of resolution maintenance, contrast recovery, and noise reduction. In patient study, the proposed method obtains a noise reduction of 44.6% ± 8.0% for the lung and the liver, while maintaining the regional mean activity in both simulated lesions and real tumors. The run time of the proposed method is only half that of the conventional IR. Significance. The proposed cycle-consistent learning from the backprojection rather than the raw PET data or an IR result enables improved reconstruction accuracy, reduced memory requirements, and fast implementation speeds for clinical whole-body PET imaging.

Convex optimization algorithms in medical image reconstruction-in the age of AI

The past decade has seen the rapid growth of model based image reconstruction (MBIR) algorithms, which are often applications or adaptations of convex optimization algorithms from the optimization community. We review some state-of-the-art algorithms that have enjoyed wide popularity in medical image reconstruction, emphasize known connections between different algorithms, and discuss practical issues such as computation and memory cost. More recently, deep learning (DL) has forayed into medical imaging, where the latest development tries to exploit the synergy between DL and MBIR to elevate the MBIR's performance. We present existing approaches and emerging trends in DL-enhanced MBIR methods, with particular attention to the underlying role of convexity and convex algorithms on network architecture. We also discuss how convexity can be employed to improve the generalizability and representation power of DL networks in general.