Search

International audience; X-ray in-line phase contrast imaging relies on the measurement of Fresnel diffraction intensity patterns due to the phase shift and the attenuation induced by the object. The recovery of phase and attenuation from one or several diffraction patterns is a nonlinear ill-posed inverse problem. In this work, we propose supervised learning approaches using mixed scale dense (MS-D) convolutional neural networks to simultaneously retrieve the phase and the attenuation from x-ray phase contrast images. This network architecture uses dilated convolutions to capture features at different image scales and densely connects all feature maps. The long range information in images becomes quickly available, and greater receptive field size can be obtained without losing resolution. This network architecture seems to account for the effect of the Fresnel operator very efficiently. We train the networks using simulated data of objects consisting of either homogeneous components, characterized by a fixed ratio of the induced refractive phase shifts and attenuation, or heterogeneous components, consisting of various materials. We also train the networks in the image domain by applying a simple initial reconstruction using the adjoint of the Fréchet derivative. We compare the results obtained with the MS-D network to reconstructions using U-Net, another popular network architecture, as well as to reconstructions using the contrast transfer function method, a direct phase and attenuation retrieval method based on linearization of the direct problem. The networks are evaluated using simulated noisy data as well as images acquired at NanoMAX (MAX IV, Lund, Sweden). In all cases, large improvements of the reconstruction errors are obtained on simulated data compared to the linearized method. Moreover, on experimental data, the networks improve the reconstruction quantitatively, improving the low-frequency behavior and the resolution.

International audience; PurposeComputed tomography (CT) is a technique of choice to image bone structure at different scales. Methods to enhance the quality of degraded reconstructions obtained from low-dose CT data have shown impressive results recently, especially in the realm of supervised deep learning. As the choice of the loss function affects the reconstruction quality, it is necessary to focus on the way neural networks evaluate the correspondence between predicted and target images during the training stage. This is even more true in the case of bone microarchitecture imaging at high spatial resolution where both the quantitative analysis of bone mineral density (BMD) and bone microstructure is essential for assessing diseases such as osteoporosis. Our aim is thus to evaluate the quality of reconstruction on key metrics for diagnosis depending on the loss function that has been used for training the neural network.MethodsWe compare and analyze volumes that are reconstructed with neural networks trained with pixelwise, structural, and adversarial loss functions or with a combination of them. We perform realistic simulations of various low-dose acquisitions of bone microarchitecture. Our comparative study is performed with metrics that have an interest regarding the diagnosis of bone diseases. We therefore focus on bone-specific metrics such as bone volume and the total volume (BV and TV), resolution, connectivity assessed with the Euler number, and quantitative analysis of BMD to evaluate the quality of reconstruction obtained with networks trained with the different loss functions.ResultsWe find that using L1norm as the pixelwise loss is the best choice compared to L2 or no pixelwise loss since it improves resolution without deteriorating other metrics. Visual Geometry Group (VGG) perceptual loss, especially when combined with an adversarial loss, allows to better retrieve topological and morphological parameters of bone microarchitecture compared to Structural SIMilarity (SSIM) index. This however leads to a decreased resolution performance. The adversarial loss enhances the reconstruction performance in terms of BMD distribution accuracy.ConclusionsIn order to retrieve the quantitative and structural characteristics of bone microarchitecture that are essential for postreconstruction diagnosis, our results suggest to use L1norm as part of the loss function. Then, trade-offs should be made depending on the application: VGG perceptual loss improves accuracy in terms of connectivity at the cost of a deteriorated resolution, and adversarial losses help better retrieve BMD distribution while significantly increasing the training time.

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.

International audience; High Resolution peripheral Quantitative CT (HR-pQCT) permits to investigate bone micro-architecture in vivo. While it is a considerable progress over standard CT for this specific application, the spatial resolution remains so far limited for the investigation of bone topology. In this work, we investigate super resolution techniques to improve 3D HR-pQCT images via semi-coupled dictionary learning based on the knowledge of high resolution micro-CT images. We propose a method to select the number of atoms and study the impact of patch sizes. To handle the anisotropy of the 3D bone structure, we propose a 2.5D strategy learning low and high resolution dictionaries in a semi-coupled way on the three different directions. The results show that this strategy is superior to the application of the method in 2D. The bone volume fraction is successfully recovered and the estimation of the connectivity is improved in most test samples. The images have a better quality compared with previously studied methods based on Total Variation (TV) regularization or a combination of TV and a double-well potential. In the future, we expect to further evaluate the method on a larger data set.

We propose a nonlinear primal–dual algorithm for the retrieval of phase shift and absorption from a single x ray in-line phase contrast, or Fresnel diffraction, image. The algorithm permits us to regularize phase and absorption separately. We demonstrate that taking into account the nonlinearity in the reconstruction improves reconstruction compared with linear methods. We also demonstrate that choosing different regularizers for absorption and phase can improve the reconstructions. The use of the total variation and its generalization in a primal–dual approach allows us to exploit the sparsity of the investigated sample. On both simulated and real datasets, the proposed nonlinear primal–dual hybrid gradient (NL-PDHG) method yields reconstructions with considerably fewer artifacts and improved the normalized mean squared error compared with its linearized version.

International audience; In this paper, we present a method of choice of an adaptative regularization parameter for data corrupted by Poisson noise based on a bilevel approach. The forward operator considered is nonlinear. The existence and unicity of the smoothed lower level problem, the differentiability properties of the constraint, and the adjoint method used to calculate the gradient of the reduced functional are studied in detail. The variance of the KL functional for Poisson noise is also investigated. The method is applied to the spectral CT inverse problem. Better reconstruction results are obtained with the bilevel method of choice than with a scalar regularization parameter.

International audience; Super resolution problems are widely discussed in medical imaging. Spatial resolution of medical images are not sufficient due to the constraints such as image acquisition time, low irradiation dose or hardware limits. To address these problems, different super resolution methods have been proposed, such as optimization or learning-based approaches. Recently, deep learning methods become a thriving technology and are developing at an exponential speed. We think it is necessary to write a review to present the current situation of deep learning in medical imaging super resolution. In this paper, we first briefly introduce deep learning methods, then present a number of important deep learning approaches to solve super resolution problems, different architectures as well as up-sampling operations will be introduced. Afterwards, we focus on the applications of deep learning methods in medical imaging super resolution problems, the challenges to overcome will be presented as well

Regularized algorithms are the state-of-the-art in computed tomography, but they are also very demanding in computer resources. In this work we test two data-fidelity formulations and some associated algorithms for the resolution of the Total-Variation regularized tomographic problem. We compare their computational cost for a mixture of Poisson and Gaussian noises. We show that a recently proposed MAP-EM algorithm outperforms the TV-regularized SIRT and the Chambolle-Pock algorithms on synthetic data for the considered noise. We illustrate this result on experimental data from transmission electron microscopy.

International audience; In this work, we study the reconstruction of blood velocity with contrast enhanced computed tomography with tomographic projections perpendicular to the main flow field direction. The inverse problem is regularized with a convection-diffusion partial differential equation and with the Navier-Stokes equation. The velocity field is reconstructed together with the density field with a first-order adjoint method. The method is validated on a simple phantom. The addition of a physical constraint with a partial differential equation for the velocity field improves the reconstruction results.

We study the properties of a regularization method for inverse problems with joint Kullback-Leibler data term and regularization when the data and the operator are corrupted by some noise. We show the convergence of the method and we obtain convergence rates for the approximate solution of the inverse problem and for the operator when it is characterized by some kernel, under the assumption that some source conditions are satisfied. Numerical results showing the effect of the noise levels on the reconstructed solution are provided for Spectral Computerized Tomography.

International audience; We study the properties of a regularization method for inverse problems corrupted by Poisson noise with Kullback-Leibler divergence as data term. The regularization parameter is chosen according to a Morozov type principle. We show that this method of choice of the parameter is well-defined. This a posteriori choice leads to a convergent regularization method. Convergences rates are obtained for this a posteriori choice of the regularization parameter when some source condition is satisfied.

A new nonlinear level-set regularization method to reconstruct the complex refractive index distribution with in-line phase contrast tomography measurements is presented under the assumption that the index is piecewise constant. The nonlinear iterative approach is based on the Frechet derivative of the intensity recorded at a single propagation distance and for several projection angles. The algorithm is successfully applied to a multi-material object for several noise levels. Better reconstruction results are achieved with a stochastic perturbation of the level-set function. This evolution corresponds to a stochastic evolution of the shape of the reconstructed regions. The reconstruction errors can be further decreased with topological derivatives. The different algorithms are tested on various multi-material objects.

International audience; Discrete tomography refers to a class of reconstruction methods adapted to discrete-valued images. Many different approaches have been investigated to address the binary case, when a two-phase object is considered. This reconstruction problem is very important in medical or material applications where it is crucial to reduce the number of projections. In this paper, we address the problem of binary image reconstruction for X-ray CT imaging from a small number of projections. We propose a TV (Total Variation) regularization approach and compare the results obtained with or without an additional box convex constraint. The schemes are applied to a simple disk image and to more complex bone cross-sections for various noise levels. The minimization of the regularization functional is performed with the state-of-the-art ADMM (Alternate Direction Minimization Method) algorithm. The methods perform equally well on a simple disk image. The additional box convex constraints improves the reconstruction results for complex structures with fine details.

International audience; Thanks to photon-counting detectors, spectral computerized tomography records energy-resolved data from which the chemical composition of a sample can be recovered. This problem, referred to as material decomposition, can be formulated as a nonlinear inverse problem. In previous work, we proposed to decompose the projection images using a regularized Gauss-Newton algorithm. To reduce further the ill-posedness of the problem, we propose here to consider equality and inequality constraints that are based on physical priors. In particular, we impose the positivity of the solutions as well the total mass in each projection image. In practice, we first decompose the projection images for each projection angle independently. Then, we reconstruct the sample slices from the decomposed projection images using a standard filtered-back projection algorithm. The constrained material decomposition problem is solved by the alternating direction method of multipliers (ADMM). We compare the proposed ADMM algorithm to the unconstrained Gauss-Newton algorithm in a numerical thorax phantom. Including constraints reduces the cross-talk between materials in both the decomposed projections and the reconstructed slices.

Over the last decade, dual-energy CT scanners have gone from prototypes to clinically available machines, and spectral photon counting CT scanners are following. They require a specific reconstruction process, consisting of two steps: material decomposition and tomographic reconstruction. Image-based methods perform reconstruction, then decomposition, while projection-based methods perform decomposition first, and then reconstruction. As an alternative, 'one-step inversion' methods have been proposed, which perform decomposition and reconstruction simultaneously. Unfortunately, one-step methods are typically slower than their two-step counterparts, and in most CT applications, reconstruction time is critical. This paper therefore proposes to compare the convergence speeds of five one-step algorithms. We adapted all these algorithms to solve the same problem: spectral photon-counting CT reconstruction from five energy bins, using a three materials decomposition basis and spatial regularization. The paper compares a Bayesian method which uses non-linear conjugate gradient for minimization (Cai et al 2013 Med. Phys. 40 111916-31), three methods based on quadratic surrogates (Long and Fessler 2014 IEEE Trans. Med. Imaging 33 1614-26, Weidinger et al 2016 Int. J. Biomed. Imaging 2016 1-15, Mechlem et al 2018 IEEE Trans. Med. Imaging 37 68-80), and a primal-dual method based on MOCCA, a modified Chambolle-Pock algorithm (Barber et al 2016 Phys. Med. Biol. 61 3784). Some of these methods have been accelerated by using μ-preconditioning, i.e. by performing all internal computations not with the actual materials the object is made of, but with carefully chosen linear combinations of those. In this paper, we also evaluated the impact of three different μ-preconditioners on convergence speed. Our experiments on simulated data revealed vast differences in the number of iterations required to reach a common image quality objective: Mechlem et al (2018 IEEE Trans. Med. Imaging 37 68-80) needed ten iterations, Cai et al (2013 Med. Phys. 40 111916-31), Long and Fessler (2014 IEEE Trans. Med. Imaging 33 1614-26) and Weidinger et al (2016 Int. J. Biomed. Imaging 2016 1-15) several hundreds, and Barber et al (2016 Phys. Med. Biol. 61 3784) several thousands. We also sum up other practical aspects, like memory footprint and the need to tune extra parameters.

In this work, we study the reconstruction of blood velocity with contrast enhanced computed tomography with a tomographic projections perpendicular to the main flow field direction. The inverse problem is regularized with a convection-diffusion partial differential equation. The velocity field is reconstructed with first order and second order adjoint methods with a receding optimal control method and tested on simple phantoms.

In this work, we investigate an inverse problem approach to 3D super-resolution/segmentation for an application to the analysis of trabecular bone micro-architecture from in vivo 3D X-ray CT images. The problem is expressed as the minimization of a functional including a data term and a prior. We consider here a regularization term combining total variation (TV) and a double-well potential to enforce the quasi-binarity of the resulting image. Three different schemes to minimize this nonconvex functional are presented and compared. The methods are applied to experimental new high-resolution peripheral quantitative CT images (voxel size $$82\,\upmu \hbox {m}$$ ) and evaluated with respect to a micro-CT image at higher spatial resolution (voxel size $$41\,\upmu \hbox {m}$$ ) considered as a ground truth. Our results show that a combination of double-well functional and TV term improves the contrast and the quality of the restoration even if the connectivity may be degraded.

While standard computed tomography (CT) data do not depend on energy, spectral computed tomography (SPCT) acquire energy-resolved data, which allows material decomposition of the object of interest. Decompo- sitions in the projection domain allow creating projection mass density (PMD) per materials. From decomposed projections, a tomographic reconstruction creates 3D material density volume. The decomposition is made pos- sible by minimizing a cost function. The variational approach is preferred since this is an ill-posed non-linear inverse problem. Moreover, noise plays a critical role when decomposing data. That is why in this paper, a new data fidelity term is used to take into account of the photonic noise. In this work two data fidelity terms were investigated: a weighted least squares (WLS) term, adapted to Gaussian noise, and the Kullback-Leibler distance (KL), adapted to Poisson noise. A regularized Gauss-Newton algorithm minimizes the cost function iteratively. Both methods decompose materials from a numerical phantom of a mouse. Soft tissues and bones are decomposed in the projection domain; then a tomographic reconstruction creates a 3D material density volume for each material. Comparing relative errors, KL is shown to outperform WLS for low photon counts, in 2D and 3D. This new method could be of particular interest when low-dose acquisitions are performed.