Skip to content

Reconstruction as an Inverse Problem ​

The function reconstruction computes an image from the measured data with the encoding operator described in Signal Model and Encoding Operators. This page explains the problem it solves, the building blocks it uses and how the reconstruction types differ.

The Inverse Problem ​

With the encoding operator , the measured samples and the unknown image satisfy  , up to noise and model errors. Computing from is an inverse problem. It is usually ill-posed: for undersampled data, many images fit the data equally well, and noise in the data can be strongly amplified. The iterative reconstructions therefore solve a regularized least-squares problem, like equation (1) in the MRIReco.jl paper but with the weights and a factor :

The regularization term expresses prior knowledge about the image, for example that it consists of regions with constant intensity or has few large wavelet coefficients, and the diagonal matrix weights the samples, as described below. Iterative Reconstruction shows how the choice of the regularization and of its weight λ affects the image. The same problem can also be solved on a GPU, as described in Run a Reconstruction on the GPU.

The direct reconstruction, "direct", doesn't solve this problem. It applies the adjoint of the encoding operator to the weighted data,  . For fully sampled Cartesian data, this is the inverse FFT, and for non-Cartesian data, it is the gridding reconstruction. With trajectories that cover a circular region of k-space, such as radial and spiral ones, the image is then set to zero outside a circle, or in 3D a sphere, with the diameter of its longest side.

Density Compensation ​

Non-Cartesian trajectories sample k-space unevenly, radial trajectories for example much more densely at the center than at the edges. The density compensation factors on the diagonal of compensate for this in the direct reconstruction. For non-Cartesian data, MRIReco.jl computes them with the iterative method of Pipe and Menon, and for Cartesian data, they are constant. samplingDensity returns their square roots, the diagonal of . The iterative reconstructions weight the data term with as well, which speeds up the convergence of the solvers. With :densityWeighting = false, they use a constant instead.

Solvers and Regularization ​

MRIReco.jl takes the solvers and the regularization terms from RegularizedLeastSquares.jl, as described in section 2.5 of the paper:

  • The parameter :solver selects the solver, for example CGNR, the conjugate gradient method applied to the normal equations, which supports an regularization, or ADMM and FISTA, which also support non-smooth terms such as the norm and total variation. These solvers only need the encoding operator, its adjoint and the normal operator.

  • The parameter :reg sets the regularization term, for example L2Regularization(λ), L1Regularization(λ) or TVRegularization(λ; shape). Each term provides its proximal map, which proximal solvers such as ADMM and FISTA use. Use a Plug-and-Play Prior shows how to replace this map by an image denoiser.

  • The parameter :sparseTrafo applies the regularization to a transformed image, except with CGNR, which ignores the transform of its term. With "Wavelet", an regularization penalizes instead of , where is the wavelet transform. Use a Custom Sparsifying Transform shows how to use other transforms and that FISTA, POGM and OptISTA handle only orthonormal ones correctly.

  • The parameter :normalizeReg scales λ. By default, λ is used as given. With MeasurementBasedNormalization, it is multiplied by the mean magnitude of the weighted data , or with FISTA, POGM and OptISTA by that of the back-projected data , so that for terms such as the norm and total variation, the same value works for data of different scale.

The examples in the paper select the solver and the regularization by name, with params[:solver] = "admm" and params[:regularization] = "TV" together with params[:λ]. The current version takes them as types and objects, params[:solver] = ADMM and params[:reg] = TVRegularization(λ; shape). It passes the solver parameters of LeastSquaresSolverParameter to the solvers that take them, under the names of their keyword arguments in RegularizedLeastSquares.jl, for example :rho instead of :ρ for the penalty parameter of ADMM and :iterationsCG instead of :iterationsInner for its inner CG iterations. Other keyword arguments of the solvers cannot be set with the dictionary, and the old names, such as :regularization, :λ and :ρ, are ignored. Reconstruction Parameters lists the keys of the dictionary.

Reconstruction Types ​

The reconstruction type, :reco, decides which data is reconstructed together and with which encoding operator. Different slices and repetitions are always reconstructed separately.

:recoReconstructed togetherEncoding operator
"direct"the data of one coil and contrastFourier encoding, only its adjoint is applied
"standard"the data of one coil and contrastFourier encoding
"multiEcho"all contrasts of one coilFourier encoding of each contrast
"multiCoil"all coils of one contrastcoil sensitivities and Fourier encoding
"multiCoilMultiEcho"all coils and contrastscoil sensitivities and Fourier encoding of each contrast
"multiCoilMultiEchoSubspace"all coils and contrastsas for "multiCoilMultiEcho", applied to the images of a temporal basis

The reconstruction types with coil sensitivities, which :senseMaps provides, combine the data of all coils into one image. The subspace reconstruction represents the images of all contrasts by a few coefficient images in a temporal basis, which :basis provides, and reconstructs these coefficient images, as shown in Subspace Reconstruction.

Noise Decorrelation ​

The noise of different receive coils is usually correlated. If a noise measurement is given with :noiseData, the reconstruction types "multiCoil", "multiCoilMultiEcho" and "multiCoilMultiEchoSubspace" estimate the noise covariance matrix from it, compute its Cholesky factorization   and apply to the data and to the coil sensitivities. This decorrelates the noise, which the least-squares problem assumes to be uncorrelated and of equal variance in all coils. The other reconstruction types ignore :noiseData. With several contrasts, "multiCoilMultiEcho" and "multiCoilMultiEchoSubspace" currently decorrelate the data wrongly, which gives a wrong image without a warning (#338).