Undersampled MR Data Research Articles

Fast and low-dose medical imaging generation empowered by hybrid deep-learning and iterative reconstruction

Fast and low-dose reconstructions of medical images are highly desired in clinical routines. We propose a hybrid deep-learning and iterative reconstruction (hybrid DL-IR) framework and apply it for fast magnetic resonance imaging (MRI), fast positron emission tomography (PET), and low-dose computed tomography (CT) image generation tasks. First, in a retrospective MRI study (6,066 cases), we demonstrate its capability of handling 3- to 10-fold under-sampled MR data, enabling organ-level coverage with only 10- to 100-s scan time; second, a low-dose CT study (142 cases) shows that our framework can successfully alleviate the noise and streak artifacts in scans performed with only 10% radiation dose (0.61 mGy); and last, a fast whole-body PET study (131 cases) allows us to faithfully reconstruct tumor-induced lesions, including small ones (<4mm), from 2- to 4-fold-accelerated PET acquisition (30-60 s/bp). This study offers a promising avenue for accurate and high-quality image reconstruction with broad clinical value.

Physics-Driven Deep Learning Methods for Fast Quantitative Magnetic Resonance Imaging: Performance improvements through integration with deep neural networks

Quantitative magnetic resonance imaging (qMRI) aims to obtain quantitative biophysical parameters based on physical models derived from MR spin magnetization evolution. This requires the acquisition of multiple MR images, resulting in very long scan times. Recently, deep learning (DL) has shown significant potential for reconstructing undersampled MR data. When applying DL to fast qMRI, physical priors can be integrated into the DL framework to improve the performance of qMRI further. This article introduces the physical models of qMRI and four ways to integrate these models into a deep neural network (NN), namely, training sample generation, contrast image prediction, loss function design, and network architecture design. The core concepts and methods for each category are presented. The associated signal processing issues regarding the generalization and reliability of physics-driven DL-based fast qMRI methods are also discussed.

Real-time radial reconstruction with domain transform manifold learning for MRI-guided radiotherapy.

MRI-guidance techniques that dynamically adapt radiation beams to follow tumor motion in real time will lead to more accurate cancer treatments and reduced collateral healthy tissue damage. The gold-standard for reconstruction of undersampled MR data is compressed sensing (CS) which is computationally slow and limits the rate that images can be available for real-timeadaptation. Once trained, neural networks can be used to accurately reconstruct raw MRI data with minimal latency. Here, we test the suitability of deep-learning-based image reconstruction for real-time tracking applications on MRI-Linacs. We use automated transform by manifold approximation (AUTOMAP), a generalized framework that maps raw MR signal to the target image domain, to rapidly reconstruct images from undersampled radial k-space data. The AUTOMAP neural network was trained to reconstruct images from a golden-angle radial acquisition, a benchmark for motion-sensitive imaging, on lung cancer patient data and generic images from ImageNet. Model training was subsequently augmented with motion-encoded k-space data derived from videos in the YouTube-8M dataset to encourage motion robustreconstruction. AUTOMAP models fine-tuned on retrospectively acquired lung cancer patient data reconstructed radial k-space with equivalent accuracy to CS but with much shorter processing times. Validation of motion-trained models with a virtual dynamic lung tumor phantom showed that the generalized motion properties learned from YouTube lead to improved target trackingaccuracy. AUTOMAP can achieve real-time, accurate reconstruction of radial data. These findings imply that neural-network-based reconstruction is potentially superior to alternative approaches for real-time image guidanceapplications.

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.

Discrete Shearlets as a Sparsifying Transform in Low-Rank Plus Sparse Decomposition for Undersampled (k, t)-Space MR Data.

The discrete shearlet transformation accurately represents the discontinuities and edges occurring in magnetic resonance imaging, providing an excellent option of a sparsifying transform. In the present paper, we examine the use of discrete shearlets over other sparsifying transforms in a low-rank plus sparse decomposition problem, denoted by . The proposed algorithm is evaluated on simulated dynamic contrast enhanced (DCE) and small bowel data. For the small bowel, eight subjects were scanned; the sequence was run first on breath-holding and subsequently on free-breathing, without changing the anatomical position of the subject. The reconstruction performance of the proposed algorithm was evaluated against k-t FOCUSS. decomposition, using discrete shearlets as sparsifying transforms, successfully separated the low-rank (background and periodic motion) from the sparse component (enhancement or bowel motility) for both DCE and small bowel data. Motion estimated from low-rank of DCE data is closer to ground truth deformations than motion estimated from L and S. Motility metrics derived from the S component of free-breathing data were not significantly different from the ones from breath-holding data up to four-fold undersampling, indicating that bowel (rapid/random) motility is isolated in S. Our work strongly supports the use of discrete shearlets as a sparsifying transform in a decomposition for undersampled MR data.

Transfer learning enhanced generative adversarial networks for multi-channel MRI reconstruction

Deep learning based generative adversarial networks (GAN) can effectively perform image reconstruction with under-sampled MR data. In general, a large number of training samples are required to improve the reconstruction performance of a certain model. However, in real clinical applications, it is difficult to obtain tens of thousands of raw patient data to train the model since saving k-space data is not in the routine clinical flow. Therefore, enhancing the generalizability of a network based on small samples is urgently needed. In this study, three novel applications were explored based on parallel imaging combined with the GAN model (PI-GAN) and transfer learning. The model was pre-trained with public Calgary brain images and then fine-tuned for use in (1) patients with tumors in our center; (2) different anatomies, including knee and liver; (3) different k-space sampling masks with acceleration factors (AFs) of 2 and 6. As for the brain tumor dataset, the transfer learning results could remove the artifacts found in PI-GAN and yield smoother brain edges. The transfer learning results for the knee and liver were superior to those of the PI-GAN model trained with its own dataset using a smaller number of training cases. However, the learning procedure converged more slowly in the knee datasets compared to the learning in the brain tumor datasets. The reconstruction performance was improved by transfer learning both in the models with AFs of 2 and 6. Of these two models, the one with AF = 2 showed better results. The results also showed that transfer learning with the pre-trained model could solve the problem of inconsistency between the training and test datasets and facilitate generalization to unseen data.

SANTIS: Sampling-Augmented Neural neTwork with Incoherent Structure for MR image reconstruction.

To develop and evaluate a novel deep learning-based reconstruction framework called SANTIS (Sampling-Augmented Neural neTwork with Incoherent Structure) for efficient MR image reconstruction with improved robustness against sampling pattern discrepancy. With a combination of data cycle-consistent adversarial network, end-to-end convolutional neural network mapping, and data fidelity enforcement for reconstructing undersampled MR data, SANTIS additionally utilizes a sampling-augmented training strategy by extensively varying undersampling patterns during training, so that the network is capable of learning various aliasing structures and thereby removing undersampling artifacts more effectively and robustly. The performance of SANTIS was demonstrated for accelerated knee imaging and liver imaging using a Cartesian trajectory and a golden-angle radial trajectory, respectively. Quantitative metrics were used to assess its performance against different references. The feasibility of SANTIS in reconstructing dynamic contrast-enhanced images was also demonstrated using transfer learning. Compared to conventional reconstruction that exploits image sparsity, SANTIS achieved consistently improved reconstruction performance (lower errors and greater image sharpness). Compared to standard learning-based methods without sampling augmentation (e.g., training with a fixed undersampling pattern), SANTIS provides comparable reconstruction performance, but significantly improved robustness, against sampling pattern discrepancy. SANTIS also achieved encouraging results for reconstructing liver images acquired at different contrast phases. By extensively varying undersampling patterns, the sampling-augmented training strategy in SANTIS can remove undersampling artifacts more robustly. The novel concept behind SANTIS can particularly be useful for improving the robustness of deep learning-based image reconstruction against discrepancy between training and inference, an important, but currently less explored, topic.

Multi-modal synergistic PET and MR reconstruction using mutually weighted quadratic priors.

PurposeTo propose a framework for synergistic reconstruction of PET‐MR and multi‐contrast MR data to improve the image quality obtained from noisy PET data and from undersampled MR data.Theory and MethodsWeighted quadratic priors were devised to preserve common boundaries between PET‐MR images while reducing noise, PET Gibbs ringing, and MR undersampling artifacts. These priors are iteratively reweighted using normalized multi‐modal Gaussian similarity kernels. Synergistic PET‐MR reconstructions were built on the PET maximum a posteriori expectation maximization algorithm and the MR regularized sensitivity encoding method. The proposed approach was compared to conventional methods, total variation, and prior‐image weighted quadratic regularization methods. Comparisons were performed on a simulated [18F]fluorodeoxyglucose‐PET and T1/T2‐weighted MR brain phantom, 2 in vivo T1/T2‐weighted MR brain datasets, and an in vivo [18F]fluorodeoxyglucose‐PET and fluid‐attenuated inversion recovery/T1‐weighted MR brain dataset.ResultsSimulations showed that synergistic reconstructions achieve the lowest quantification errors for all image modalities compared to conventional, total variation, and weighted quadratic methods. Whereas total variation regularization preserved modality‐unique features, this method failed to recover PET details and was not able to reduce MR artifacts compared to our proposed method. For in vivo MR data, our method maintained similar image quality for 3× and 14× accelerated data. Reconstruction of the PET‐MR dataset also demonstrated improved performance of our method compared to the conventional independent methods in terms of reduced Gibbs and undersampling artifacts.ConclusionThe proposed methodology offers a robust multi‐modal synergistic image reconstruction framework that can be readily built on existing established algorithms.

Respiratory motion compensation for simultaneous PET/MR based on highly undersampled MR data.

Positron emission tomography (PET) of the thorax region is impaired by respiratory patient motion. To account for motion, the authors propose a new method for PET/magnetic resonance (MR) respiratory motion compensation (MoCo), which uses highly undersampled MR data with acquisition times as short as 1 min/bed. The proposed PET/MR MoCo method (4D jMoCo PET) uses radial MR data to estimate the respiratory patient motion employing MR joint motion estimation and image reconstruction with temporal median filtering. Resulting motion vector fields are incorporated into the system matrix of the PET reconstruction. The proposed approach is evaluated for the thorax region utilizing a PET/MR simulation with 1 min MR acquisition time and simultaneous PET/MR measurements of six patients with MR acquisition times of 1 and 5 min and radial undersampling factors of 11.2 and 2.2, respectively. Reconstruction results are compared to 3D PET, 4D gated PET and a standard MoCo method (4D sMoCo PET), which performs iterative image reconstruction and motion estimation sequentially. Quantitative analysis comprises the parameters SUVmean, SUVmax, full width at half-maximum/lesion volume, contrast and signal-to-noise ratio. For simulated PET data, our quantitative analysis shows that the proposed 4D jMoCo PET approach with temporal filtering achieves the best quantification accuracy of all tested reconstruction methods with a mean absolute deviation of 2.3% when compared to the ground truth. For measured PET patient data, the mean absolute deviation of 4D jMoCo PET using a 1 min MR acquisition for motion estimation is 2.1% relative to the 5 min MR acquisition. This demonstrates a robust behavior even in case of strong undersampling at MR acquisition times as short as 1 min. In contrast, 4D sMoCo PET shows considerable reduction of quantification accuracy for the 1 min MR acquisition time. Relative to 3D PET, the proposed 4D jMoCo PET approach with temporal filtering yields an average increase of SUVmean, SUVmax, and contrast of 29.9% and 13.8% for simulated and measured PET data, respectively. Employing artifact-robust motion estimation enables PET/MR respiratory MoCo with MR acquisition times as short as 1 min/bed improving PET image quality and quantification accuracy.

A statistical method for characterizing the noise in nonlinearly reconstructed images from undersampled MR data: The POCS example

The projection-onto-convex-sets (POCS) algorithm is a powerful tool for reconstructing high-resolution images from undersampled k-space data. It is a nonlinear iterative method that attempts to estimate values for missing data. The convergence of the algorithm and its other deterministic properties are well established, but relatively little is known about how noise in the source data influences noise in the final reconstructed image. In this paper, we present an experimental treatment of the statistical properties in POCS and investigate 12 stochastic models for its noise distribution beside its nonlinear point spread functions. Statistical results show that as the ratio of the missing k-space data increases, the noise distribution in POCS images is no longer Rayleigh as with conventional linear Fourier reconstruction. Instead, the probability density function for the noise is well approximated by a lognormal distribution. For small missing data ratios, however, the noise remains Rayleigh distributed. Preliminary results show that in the presence of noise, POCS images are often dominated by POCS-enhanced noise rather than POCS-induced artifacts. Implicit in this work is the presentation of a general statistical method that can be used to assess the noise properties in other nonlinear reconstruction algorithms.

Self-calibrating Coil Sensitivity Profiles for Parallel Imaging Based on Anisotropic Diffusion

Calibration of the spatial sensitivity functions of coil arrays is a crucial element in parallel magnetic resonance imaging (PMRI).In order to overcome the sensitivity miscalibration errors introduced by pre-scan calibration method and decrease the total examination time , a new self-calibrating method for estimating the coil sensitivity profiles was proposed i n th e paper . F rom a fully - sampled central region of a variable-density k -space acquisition, t he sensitivity calibration images were directly extracted , and then were adaptively smoothed by a new anisotropic diffusion scheme to extract encoding effects of coil sensitivities. When the estimated coil sensitivity profiles were applied to reconstruct the full Field-Of-View image from the uniformly under-sampled MR data when under-sampl e rate was greater than 2 , the experimental results showed that the quality of reconstruct ion image was evidently improv ed. Therefore, t h is self-calibrating method for calculating sensitivity profiles is suitable for much faster imaging in parallel magnetic imaging. Meanwhile, the paper suggest that the suitable number of central k-space lines for generating the internal reference images should be chosen between 8 to 20.