Ground-truth-free deep learning for artefacts reduction in 2D radial cardiac cine MRI using a synthetically generated dataset.

In this work, we consider the task of image reconstruction in 2D radial cardiac cine MRI using deep learning (DL)-based regularization. As the regularization is achieved by employing an image-prior predicted by a pre-trained convolutional neural network (CNN), the quality of the image-prior is of essential importance. The achievable performance of any DL-based method is limited by the amount and the quality of the available training data. For fast dynamic processes, obtaining good-quality MR data is challenging because of technical and physiological reasons. In this work, we try to overcome these problems by a transfer-learning approach which is motivated by a previously presented DL-method (XT,YT U-Net). There, instead of training the network on the whole 2D dynamic images, it is trained on 2D spatio-temporal profiles (xt,yt-slices) which show the temporal changes of the imaged object. Therefore, for the training and test data, it is more important that their spatio-temporal profiles share similar local features rather than being images of the same anatomy. This allows us to equip arbitrary data with simulated motion that resembles the cardiac motion and use it as training data. By doing so, it is possible to train a CNN which is applicable to cardiac cine MR data without using ground-truth cine MR images for training. We demonstrate that combining XT,YT U-Net with the proposed transfer-learning strategy delivers comparable performance to CNNs trained on cardiac cine MR images and in some cases even qualitatively surpasses these. Additionally, the transfer-learning strategy was investigated for a 2D and 3D U-Net. The images processed by the the CNNs were used as image-priors in the CNN-regularized iterative reconstruction. The XT,YT U-Net yielded visibly better results than the 2D U-Net and slightly better results than the 3D U-Net when used in combination with the presented transfer learning-strategy.

Neural networks-based regularization for large-scale medical image reconstruction.

In this paper we present a generalized Deep Learning-based approach for solving ill-posed large-scale inverse problems occuring in medical image reconstruction. Recently, Deep Learning methods using iterative neural networks (NNs) and cascaded NNs have been reported to achieve state-of-the-art results with respect to various quantitative quality measures as PSNR, NRMSE and SSIM across different imaging modalities. However, the fact that these approaches employ the application of the forward and adjoint operators repeatedly in the network architecture requires the network to process the whole images or volumes at once, which for some applications is computationally infeasible. In this work, we follow a different reconstruction strategy by strictly separating the application of the NN, the regularization of the solution and the consistency with the measured data. The regularization is given in the form of an image prior obtained by the output of a previously trained NN which is used in a Tikhonov regularization framework. By doing so, more complex and sophisticated network architectures can be used for the removal of the artefacts or noise than it is usually the case in iterative NNs. Due to the large scale of the considered problems and the resulting computational complexity of the employed networks, the priors are obtained by processing the images or volumes as patches or slices. We evaluated the method for the cases of 3D cone-beam low dose CT and undersampled 2D radial cine MRI and compared it to a total variation-minimization-based reconstruction algorithm as well as to a method with regularization based on learned overcomplete dictionaries. The proposed method outperformed all the reported methods with respect to all chosen quantitative measures and further accelerates the regularization step in the reconstruction by several orders of magnitude.

Determination of contrast-detail curves in mammography image quality assessment by a parametric model observer.

A novel approach is proposed for the determination of contrast-detail curves in mammography image quality assessment. The approach is compared with current practice using virtual mammography. A binary parametric model observer is applied to images of the CDMAM phantom. The observer accounts for the simple disc shaped objects in the phantom and is applied separately to each cell of the phantom. For each of these applications, the area under the ROC curve (AUC) of the model observer is determined. The different AUCs, calculated from different applications of the parametric model observer, are then combined to a single contrast-detail curve quantifying the ability of the observer to detect details in the images. Virtual mammography is developed as a tool to simulate X-ray images of single CDMAM cells and to quantitatively assess the approach in comparison with current practice. It is shown that the proposed approach can lead to similar contrast-detail curves as current practice. The precision of the estimated contrast-detail curves is increased, i.e. using 5 images yields about the same precision for the proposed approach as 16 images when applying current practice. We conclude that contrast-detail curves in mammography image quality assessment can also be determined through the AUC of a binary parametric model observer. Since the proposed approach has higher precision than current practice, it is a promising candidate for contrast-detail analysis in mammography image quality assessment.

A novel methodology for personalized simulations of ventricular hemodynamics from noninvasive imaging data.

Current state-of-the-art imaging techniques can provide quantitative information to characterize ventricular function within the limits of the spatiotemporal resolution achievable in a realistic acquisition time. These imaging data can be used to personalize computer models, which in turn can help treatment planning by quantifying biomarkers that cannot be directly imaged, such as flow energy, shear stress and pressure gradients. To date, computer models have typically relied on invasive pressure measurements to be made patient-specific. When these data are not available, the scope and validity of the models are limited. To address this problem, we propose a new methodology for modeling patient-specific hemodynamics based exclusively on noninvasive velocity and anatomical data from 3D+t echocardiography or Magnetic Resonance Imaging (MRI). Numerical simulations of the cardiac cycle are driven by the image-derived velocities prescribed at the model boundaries using a penalty method that recovers a physical solution by minimizing the energy imparted to the system. This numerical approach circumvents the mathematical challenges due to the poor conditioning that arises from the imposition of boundary conditions on velocity only. We demonstrate that through this technique we are able to reconstruct given flow fields using Dirichlet only conditions. We also perform a sensitivity analysis to investigate the accuracy of this approach for different images with varying spatiotemporal resolution. Finally, we examine the influence of noise on the computed result, showing robustness to unbiased noise with an average error in the simulated velocity approximately 7% for a typical voxel size of 2mm(3) and temporal resolution of 30ms. The methodology is eventually applied to a patient case to highlight the potential for a direct clinical translation.

Whole left ventricular functional assessment from two minutes free breathing multi-slice CINE acquisition.

Two major challenges in cardiovascular MRI are long scan times due to slow MR acquisition and motion artefacts due to respiratory motion. Recently, a Motion Corrected-Compressed Sensing (MC-CS) technique has been proposed for free breathing 2D dynamic cardiac MRI that addresses these challenges by simultaneously accelerating MR acquisition and correcting for any arbitrary motion in a compressed sensing reconstruction. In this work, the MC-CS framework is combined with parallel imaging for further acceleration, and is termed Motion Corrected Sparse SENSE (MC-SS). Validation of the MC-SS framework is demonstrated in eight volunteers and three patients for left ventricular functional assessment and results are compared with the breath-hold acquisitions as reference. A non-significant difference (P > 0.05) was observed in the volumetric functional measurements (end diastolic volume, end systolic volume, ejection fraction) and myocardial border sharpness values obtained with the proposed and gold standard methods. The proposed method achieves whole heart multi-slice coverage in 2 min under free breathing acquisition eliminating the time needed between breath-holds for instructions and recovery. This results in two-fold speed up of the total acquisition time in comparison to the breath-hold acquisition.

Towards a fast and efficient approach for modelling the patient-specific ventricular haemodynamics.

Computer modelling of the heart has emerged over the past decade as a powerful technique to explore the cardiovascular pathophysiology and inform clinical diagnosis. The current state-of-the-art in biophysical modelling requires a wealth of, potentially invasive, clinical data for the parametrisation and validation of the models, a process that is still too long and complex to be compatible with the clinical decision-making time. Therefore, there remains a need for models that can be quickly customised to reconstruct physical processes difficult to measure directly in patients. In this paper, we propose a less resource-intensive approach to modelling, whereby computational fluid-dynamics (CFD) models are constrained exclusively by boundary motion derived from imaging data through a validated wall tracking algorithm. These models are generated and parametrised based solely on ultrasound data, whose acquisition is fast, inexpensive and routine in all patients. To maximise the time and computational efficiency, a semi-automated pipeline is embedded in an image processing workflow to personalise the models. Applying this approach to two patient cases, we demonstrate this tool can be directly used in the clinic to interpret and complement the available clinical data by providing a quantitative indication of clinical markers that cannot be easily derived from imaging, such as pressure gradients and the flow energy.

Improved UTE-based attenuation correction for cranial PET-MR using dynamic magnetic field monitoring.

PURPOSE: Ultrashort echo time (UTE) MRI has been proposed as a way to produce segmented attenuation maps for PET, as it provides contrast between bone, air, and soft tissue. However, UTE sequences require samples to be acquired during rapidly changing gradient fields, which makes the resulting images prone to eddy current artifacts. In this work it is demonstrated that this can lead to misclassification of tissues in segmented attenuation maps (AC maps) and that these effects can be corrected for by measuring the true k-space trajectories using a magnetic field camera. METHODS: The k-space trajectories during a dual echo UTE sequence were measured using a dynamic magnetic field camera. UTE images were reconstructed using nominal trajectories and again using the measured trajectories. A numerical phantom was used to demonstrate the effect of reconstructing with incorrect trajectories. Images of an ovine leg phantom were reconstructed and segmented and the resulting attenuation maps were compared to a segmented map derived from a CT scan of the same phantom, using the Dice similarity measure. The feasibility of the proposed method was demonstrated in in vivo cranial imaging in five healthy volunteers. Simulated PET data were generated for one volunteer to show the impact of misclassifications on the PET reconstruction. RESULTS: Images of the numerical phantom exhibited blurring and edge artifacts on the bone-tissue and air-tissue interfaces when nominal k-space trajectories were used, leading to misclassification of soft tissue as bone and misclassification of bone as air. Images of the tissue phantom and the in vivo cranial images exhibited the same artifacts. The artifacts were greatly reduced when the measured trajectories were used. For the tissue phantom, the Dice coefficient for bone in MR relative to CT was 0.616 using the nominal trajectories and 0.814 using the measured trajectories. The Dice coefficients for soft tissue were 0.933 and 0.934 for the nominal and measured cases, respectively. For air the corresponding figures were 0.991 and 0.993. Compared to an unattenuated reference image, the mean error in simulated PET uptake in the brain was 9.16% when AC maps derived from nominal trajectories was used, with errors in the SUV max for simulated lesions in the range of 7.17%-12.19%. Corresponding figures when AC maps derived from measured trajectories were used were 0.34% (mean error) and -0.21% to +1.81% (lesions). CONCLUSIONS: Eddy current artifacts in UTE imaging can be corrected for by measuring the true k-space trajectories during a calibration scan and using them in subsequent image reconstructions. This improves the accuracy of segmented PET attenuation maps derived from UTE sequences and subsequent PET reconstruction.

A technical assessment of pulse wave velocity algorithms applied to non-invasive arterial waveforms.

Non-invasive assessment of arterial stiffness through pulse wave velocity (PWV) analysis is becoming common clinical practice. However, the effects of measurement noise, temporal resolution and similarity of the two waveforms used for PWV calculation upon accuracy and variability are unknown. We studied these effects upon PWV estimates given by foot-to-foot, least squared difference, and cross-correlation algorithms. We assessed accuracy using numerically generated blood pressure and flow waveforms for which the theoretical PWV was known to compare with the algorithm estimates. We assessed variability using clinical measurements in 28 human subjects. Wave shape similarity was quantified using a cross correlation-coefficient (CCCoefficient), which decreases with increasing distance between waveform measurements sites. Based on our results, we propose the following criteria to identify the most accurate and least variable algorithm given the noise, resolution and CCCoefficient of the measured waveforms. (1) Use foot-to-foot when the noise-to-signal ratio ≤10%, and/or temporal resolution ≥100 Hz. Otherwise (2) use a least squares differencing method applied to the systolic upstroke.

Benchmarking framework for myocardial tracking and deformation algorithms: an open access database.

In this paper we present a benchmarking framework for the validation of cardiac motion analysis algorithms. The reported methods are the response to an open challenge that was issued to the medical imaging community through a MICCAI workshop. The database included magnetic resonance (MR) and 3D ultrasound (3DUS) datasets from a dynamic phantom and 15 healthy volunteers. Participants processed 3D tagged MR datasets (3DTAG), cine steady state free precession MR datasets (SSFP) and 3DUS datasets, amounting to 1158 image volumes. Ground-truth for motion tracking was based on 12 landmarks (4 walls at 3 ventricular levels). They were manually tracked by two observers in the 3DTAG data over the whole cardiac cycle, using an in-house application with 4D visualization capabilities. The median of the inter-observer variability was computed for the phantom dataset (0.77 mm) and for the volunteer datasets (0.84 mm). The ground-truth was registered to 3DUS coordinates using a point based similarity transform. Four institutions responded to the challenge by providing motion estimates for the data: Fraunhofer MEVIS (MEVIS), Bremen, Germany; Imperial College London - University College London (IUCL), UK; Universitat Pompeu Fabra (UPF), Barcelona, Spain; Inria-Asclepios project (INRIA), France. Details on the implementation and evaluation of the four methodologies are presented in this manuscript. The manually tracked landmarks were used to evaluate tracking accuracy of all methodologies. For 3DTAG, median values were computed over all time frames for the phantom dataset (MEVIS=1.20mm, IUCL=0.73 mm, UPF=1.10mm, INRIA=1.09 mm) and for the volunteer datasets (MEVIS=1.33 mm, IUCL=1.52 mm, UPF=1.09 mm, INRIA=1.32 mm). For 3DUS, median values were computed at end diastole and end systole for the phantom dataset (MEVIS=4.40 mm, UPF=3.48 mm, INRIA=4.78 mm) and for the volunteer datasets (MEVIS=3.51 mm, UPF=3.71 mm, INRIA=4.07 mm). For SSFP, median values were computed at end diastole and end systole for the phantom dataset(UPF=6.18 mm, INRIA=3.93 mm) and for the volunteer datasets (UPF=3.09 mm, INRIA=4.78 mm). Finally, strain curves were generated and qualitatively compared. Good agreement was found between the different modalities and methodologies, except for radial strain that showed a high variability in cases of lower image quality.

The effect of regularization in motion compensated PET image reconstruction: a realistic numerical 4D simulation study.

Following continuous improvement in PET spatial resolution, respiratory motion correction has become an important task. Two of the most common approaches that utilize all detected PET events to motion-correct PET data are the reconstruct-transform-average method (RTA) and motion-compensated image reconstruction (MCIR). In RTA, separate images are reconstructed for each respiratory frame, subsequently transformed to one reference frame and finally averaged to produce a motion-corrected image. In MCIR, the projection data from all frames are reconstructed by including motion information in the system matrix so that a motion-corrected image is reconstructed directly. Previous theoretical analyses have explained why MCIR is expected to outperform RTA. It has been suggested that MCIR creates less noise than RTA because the images for each separate respiratory frame will be severely affected by noise. However, recent investigations have shown that in the unregularized case RTA images can have fewer noise artefacts, while MCIR images are more quantitatively accurate but have the common salt-and-pepper noise. In this paper, we perform a realistic numerical 4D simulation study to compare the advantages gained by including regularization within reconstruction for RTA and MCIR, in particular using the median-root-prior incorporated in the ordered subsets maximum a posteriori one-step-late algorithm. In this investigation we have demonstrated that MCIR with proper regularization parameters reconstructs lesions with less bias and root mean square error and similar CNR and standard deviation to regularized RTA. This finding is reproducible for a variety of noise levels (25, 50, 100 million counts), lesion sizes (8 mm, 14 mm diameter) and iterations. Nevertheless, regularized RTA can also be a practical solution for motion compensation as a proper level of regularization reduces both bias and mean square error.

Group sparse reconstruction using intensity-based clustering.

Compressed sensing has been of great interest to speed up the acquisition of MR images. The k-t group sparse (k-t GS) method has recently been introduced for dynamic MR images to exploit not just the sparsity, as in compressed sensing, but also the spatial group structure in the sparse representation. k-t GS achieves higher acceleration factors compared to the conventional compressed sensing method. However, it assumes a spatial structure in the sparse representation and it requires a time consuming hard-thresholding reconstruction scheme. In this work, we propose to modify k-t GS by incorporating prior information about the sorted intensity of the signal in the sparse representation, for a more general and robust group assignment. This approach is referred to as group sparse reconstruction using intensity-based clustering. The feasibility of the proposed method is demonstrated for static 3D hyperpolarized lung images and applications with both dynamic and intensity changes, such as 2D cine and perfusion cardiac MRI, with retrospective undersampling. For all reported acceleration factors the proposed method outperforms the original compressed sensing method. Improved reconstruction over k-t GS method is demonstrated when k-t GS assumptions are not satisfied. The proposed method was also applied to cardiac cine images with a prospective sevenfold acceleration, outperforming the standard compressed sensing reconstruction.

Patient-specific respiratory models using dynamic 3D MRI: preliminary volunteer results.

Organ and tumour motion has a significant impact on the planning and delivery of radiotherapy treatment. At present imaging modality such as four-dimensional computer tomography (4DCT) cannot be used to measure the variability of motion between different respiratory cycles. To create reliable motion models, one needs to acquire volumetric data sets of the lungs with sufficient sampling of the breathing cycle. In this paper we investigate the use of highly parallel MRI to acquire such data. A 32 channel coil in conjunction with a balanced SSFP sequence and a SENSE factor of 6 were used to acquire volumetric data sets in five healthy volunteers. The acquisition was repeated for seven series of different breathing patterns. The data acquired was of sufficient spatial resolution (5 × 5 × 5 mm(3)) and image quality to carry out automated non-rigid registration. The acquisition rate (c.a. 2 volumes per second) allowed for a meaningful sampling of the different respiratory curves that were automatically obtained from the skin surface motion. This acquisition technique should provide images of high enough quality to create statistical respiratory models.

Respiratory motion models: a review.

The problem of respiratory motion has proved a serious obstacle in developing techniques to acquire images or guide interventions in abdominal and thoracic organs. Motion models offer a possible solution to these problems, and as a result the field of respiratory motion modelling has become an active one over the past 15 years. A motion model can be defined as a process that takes some surrogate data as input and produces a motion estimate as output. Many techniques have been proposed in the literature, differing in the data used to form the models, the type of model employed, how this model is computed, the type of surrogate data used as input to the model in order to make motion estimates and what form this output should take. In addition, a wide range of different application areas have been proposed. In this paper we summarise the state of the art in this important field and in the process highlight the key papers that have driven its advance. The intention is that this will serve as a timely review and comparison of the different techniques proposed to date and as a basis to inform future research in this area.

Nonrigid motion modeling of the liver from 3-D undersampled self-gated golden-radial phase encoded MRI.

Magnetic resonance imaging (MRI) has been commonly used for guiding and planning image guided interventions since it provides excellent soft tissue visualization of anatomy and allows motion modeling to predict the position of target tissues during the procedure. However, MRI-based motion modeling remains challenging due to the difficulty of acquiring multiple motion-free 3-D respiratory phases with adequate contrast and spatial resolution. Here, we propose a novel retrospective respiratory gating scheme from a 3-D undersampled high-resolution MRI acquisition combined with fast and robust image registrations to model the nonrigid deformation of the liver. The acquisition takes advantage of the recently introduced golden-radial phase encoding (G-RPE) trajectory. G-RPE is self-gated, i.e., the respiratory signal can be derived from the acquired data itself, and allows retrospective reconstructions of multiple respiratory phases at any arbitrary respiratory position. Nonrigid motion modeling is applied to predict the liver deformation of an average breathing cycle. The proposed approach was validated on 10 healthy volunteers. Motion model accuracy was assessed using similarity-, surface-, and landmark-based validation methods, demonstrating precise model predictions with an overall target registration error of TRE = 1.70 ± 0.94 mm which is within the range of the acquired resolution.

Thoracic respiratory motion estimation from MRI using a statistical model and a 2-D image navigator.

Respiratory motion models have potential application for estimating and correcting the effects of motion in a wide range of applications, for example in PET-MR imaging. Given that motion cycles caused by breathing are only approximately repeatable, an important quality of such models is their ability to capture and estimate the intra- and inter-cycle variability of the motion. In this paper we propose and describe a technique for free-form nonrigid respiratory motion correction in the thorax. Our model is based on a principal component analysis of the motion states encountered during different breathing patterns, and is formed from motion estimates made from dynamic 3-D MRI data. We apply our model using a data-driven technique based on a 2-D MRI image navigator. Unlike most previously reported work in the literature, our approach is able to capture both intra- and inter-cycle motion variability. In addition, the 2-D image navigator can be used to estimate how applicable the current motion model is, and hence report when more imaging data is required to update the model. We also use the motion model to decide on the best positioning for the image navigator. We validate our approach using MRI data acquired from 10 volunteers and demonstrate improvements of up to 40.5% over other reported motion modelling approaches, which corresponds to 61% of the overall respiratory motion present. Finally we demonstrate one potential application of our technique: MRI-based motion correction of real-time PET data for simultaneous PET-MRI acquisition.

Fast generation of 4D PET-MR data from real dynamic MR acquisitions.

We have implemented and evaluated a framework for simulating simultaneous dynamic PET-MR data using the anatomic and dynamic information from real MR acquisitions. PET radiotracer distribution is simulated by assigning typical FDG uptake values to segmented MR images with manually inserted additional virtual lesions. PET projection data and images are simulated using analytic forward projections (including attenuation and Poisson statistics) implemented within the image reconstruction package STIR. PET image reconstructions are also performed with STIR. The simulation is validated with numerical simulation based on Monte Carlo (GATE) which uses more accurate physical modelling, but has 150× slower computation time compared to the analytic method for ten respiratory positions and is 7000× slower when performing multiple realizations. Results are validated in terms of region of interest mean values and coefficients of variation for 65 million coincidences including scattered events. Although some discrepancy is observed, agreement between the two different simulation methods is good given the statistical noise in the data. In particular, the percentage difference of the mean values is 3.1% for tissue, 17% for the lungs and 18% for a small lesion. The utility of the procedure is demonstrated by simulating realistic PET-MR datasets from multiple volunteers with different breathing patterns. The usefulness of the toolkit will be shown for performance investigations of the reconstruction, motion correction and attenuation correction algorithms for dynamic PET-MR data.

The challenges of integrating molecular imaging into the optimization of cancer therapy.

We review novel, in vivo and tissue-based imaging technologies that monitor and optimize cancer therapeutics. Recent advances in cancer treatment centre around the development of targeted therapies and personalisation of treatment regimes to individual tumour characteristics. However, clinical outcomes have not improved as expected. Further development of the use of molecular imaging to predict or assess treatment response must address spatial heterogeneity of cancer within the body. A combination of different imaging modalities should be used to relate the effect of the drug to dosing regimen or effective drug concentration at the local site of action. Molecular imaging provides a functional and dynamic read-out of cancer therapeutics, from nanometre to whole body scale. At the whole body scale, an increase in the sensitivity and specificity of the imaging probe is required to localise (micro)metastatic foci and/or residual disease that are currently below the limit of detection. The use of image-guided endoscopic biopsy can produce tumour cells or tissues for nanoscopic analysis in a relatively patient-compliant manner, thereby linking clinical imaging to a more precise assessment of molecular mechanisms. This multimodality imaging approach (in combination with genetics/genomic information) could be used to bridge the gap between our knowledge of mechanisms underlying the processes of metastasis, tumour dormancy and routine clinical practice. Treatment regimes could therefore be individually tailored both at diagnosis and throughout treatment, through monitoring of drug pharmacodynamics providing an early read-out of response or resistance.

k-t Group sparse: a method for accelerating dynamic MRI.

Compressed sensing (CS) is a data-reduction technique that has been applied to speed up the acquisition in MRI. However, the use of this technique in dynamic MR applications has been limited in terms of the maximum achievable reduction factor. In general, noise-like artefacts and bad temporal fidelity are visible in standard CS MRI reconstructions when high reduction factors are used. To increase the maximum achievable reduction factor, additional or prior information can be incorporated in the CS reconstruction. Here, a novel CS reconstruction method is proposed that exploits the structure within the sparse representation of a signal by enforcing the support components to be in the form of groups. These groups act like a constraint in the reconstruction. The information about the support region can be easily obtained from training data in dynamic MRI acquisitions. The proposed approach was tested in two-dimensional cardiac cine MRI with both downsampled and undersampled data. Results show that higher acceleration factors (up to 9-fold), with improved spatial and temporal quality, can be obtained with the proposed approach in comparison to the standard CS reconstructions.

A computationally efficient OMP-based compressed sensing reconstruction for dynamic MRI.

Compressed sensing (CS) methods in MRI are computationally intensive. Thus, designing novel CS algorithms that can perform faster reconstructions is crucial for everyday applications. We propose a computationally efficient orthogonal matching pursuit (OMP)-based reconstruction, specifically suited to cardiac MR data. According to the energy distribution of a y-f space obtained from a sliding window reconstruction, we label the y-f space as static or dynamic. For static y-f space images, a computationally efficient masked OMP reconstruction is performed, whereas for dynamic y-f space images, standard OMP reconstruction is used. The proposed method was tested on a dynamic numerical phantom and two cardiac MR datasets. Depending on the field of view composition of the imaging data, compared to the standard OMP method, reconstruction speedup factors ranging from 1.5 to 2.5 are achieved.