Image reconstruction by domain-transform manifold learning

Bo ZhuJeremiah Z. LiuBruce R. RosenMatthew S. Rosen

article2017Nature1,705 citations

Introduces AUTOMAP, a unified deep learning framework that learns direct transforms from raw sensor data to images across diverse acquisition strategies, eliminating ad hoc reconstruction pipelines while reducing noise and artifacts.

Listen

Modern imaging systems across medicine, astronomy, and materials science rely on image reconstruction to transform raw sensor measurements into viewable images. Standard reconstruction approaches typically depend on rigid, handcrafted signal processing pipelines that are tailored to specific hardware and often require manual parameter tuning. These conventional methods are especially vulnerable when scans are accelerated or radiation doses are lowered, conditions that reduce the signal-to-noise ratio and introduce severe image artifacts.

The article demonstrates and evaluates a unified deep learning framework named AUTOMAP (Automated Transform by Manifold Approximation), designed to automatically learn the end-to-end mathematical mapping from raw sensor data directly to output images across diverse acquisition strategies using a single neural network architecture.

To demonstrate this capability, the researchers implemented a deep neural network consisting of fully connected layers coupled with a sparse convolutional autoencoder. They evaluated the framework on human brain magnetic resonance imaging (MRI) benchmarks across four demanding acquisition scenarios: Radon projection data, spiral non-Cartesian sampling, 40% Poisson-disc undersampling, and hardware-misaligned sampling. Notably, for most tasks the model was trained entirely on generic photographs of natural scenes from ImageNet rather than medical images, testing its ability to generalize to unseen biological structures under varying levels of noise.

The analysis produced several key findings. First, AUTOMAP successfully learned accurate reconstruction mappings across all four acquisition strategies without altering network hyperparameters or architectures between tasks. Second, the framework demonstrated superior immunity to noise and sampling artifacts compared with standard techniques, eliminating white noise amplification, ringing, compressed sensing distortions, and aliasing. Third, internal layer evaluations revealed that training on structured image data naturally induced sparse hidden-layer activations and organized spatial weight correlations, explaining its robust noise suppression. Finally, by incorporating synthetic phase modulations during training, the network accurately reconstructed both complex magnitude and phase images from in vivo MRI scans.

These findings indicate that image reconstruction can be unified into a single data-driven framework, eliminating the need for custom, human-engineered processing chains. By providing robust reconstruction at low signal-to-noise levels, this approach enables faster scan times and lower radiation doses in clinical settings without degrading diagnostic quality. It also allows standard public photographic and medical repositories to serve as effective training data.

Organizations and research teams developing imaging systems should consider piloting data-driven reconstruction models to enhance existing hardware performance and explore novel, non-traditional sensor sampling patterns. Next steps should focus on scaling the framework to higher image resolutions and testing performance across other clinical modalities such as low-dose computed tomography, ultrasound, and optical coherence tomography.

Readers should note that the evaluations were primarily conducted at an image matrix size of 128x128 pixels, and fully connected deep learning architectures can impose significant memory and computational demands when scaling to high-resolution volumetric datasets. While confidence in the demonstrated noise immunity and cross-domain generalization is high under the tested conditions, clinical deployment will require validation on full-scale clinical workflows and larger patient cohorts.

Harvard University Massachusetts General Hospital image acquisition low-dimensional manifolds network architecture domain-transform manifold learning image reconstruction MRI reconstruction

arXiv: 1704.08841 MattRosenLab/AUTOMAP

Paper: Deep Convolutional Neural Network for Inverse Problems in Imaging, Kyong Hwan Jin et al. (2016). This foundational paper establishes the framework of using deep convolutional neural networks to solve imaging inverse problems, providing essential context for deep learned reconstruction.
Paper: Understanding deep image representations by inverting them, Aravindh Mahendran et al. (2014). Understanding how deep visual representations can be inverted to reconstruct images provides foundational insight into learning inverse mappings from encoded representations.
Paper: Image Super-Resolution Using Deep Convolutional Networks, Chao Dong et al. (2014). This seminal work demonstrates end-to-end mapping from degraded measurements to high-resolution images using neural networks, underpinning deep learning-based image reconstruction.
Paper: Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections, Xiao-Jiao Mao et al. (2016). This paper establishes the use of deep encoder-decoder networks with skip connections for visual restoration, informing architecture designs in deep learned inverse problem solvers.

Paper: Learning a Variational Network for Reconstruction of Accelerated MRI Data, Kerstin Hammernik et al. (2017). This work extends deep learned MRI reconstruction by explicitly incorporating physical variational network formulations to accelerate multi-channel MRI data acquisition.
Paper: An overview of deep learning in medical imaging focusing on MRI, Alexander Selvikvåg Lundervold et al. (2018). This comprehensive overview reviews the broader landscape and clinical integration of deep learning across magnetic resonance imaging workflows, contextualizing manifold-learning reconstruction.
Paper: Noise2Noise: Learning Image Restoration without Clean Data, Jaakko Lehtinen et al. (2018). This work advances learned image restoration and undersampled MRI reconstruction by demonstrating how neural networks can train directly on corrupted paired data without requiring clean ground truth.
Paper: Diffusion Posterior Sampling for General Noisy Inverse Problems, Hyungjin Chung et al. (2022). This study pushes learned inverse problem solvers further by using generative diffusion models and posterior sampling to handle general noisy and nonlinear imaging transforms.
Paper: Deep Image Prior, Dmitry Ulyanov et al. (2017). This paper investigates how network architectures implicitly regularize inverse problems, offering a complementary perspective to supervised manifold transform learning.

Cover for Image reconstruction by domain-transform manifold learning

Abstract

Image reconstruction plays a critical role in the implementation of all contemporary imaging modalities across the physical and life sciences including optical, MRI, CT, PET, and radio astronomy. During an image acquisition, the sensor encodes an intermediate representation of an object in the sensor domain, which is subsequently reconstructed into an image by an inversion of the encoding function. Image reconstruction is challenging because analytic knowledge of the inverse transform may not exist a priori, especially in the presence of sensor non-idealities and noise. Thus, the standard reconstruction approach involves approximating the inverse function with multiple ad hoc stages in a signal processing chain whose composition depends on the details of each acquisition strategy, and often requires expert parameter tuning to optimize reconstruction performance. We present here a unified framework for image reconstruction, AUtomated TransfOrm by Manifold APproximation (AUTOMAP), which recasts image reconstruction as a data-driven, supervised learning task that allows a mapping between sensor and image domain to emerge from an appropriate corpus of training data. We implement AUTOMAP with a deep neural network and exhibit its flexibility in learning reconstruction transforms for a variety of MRI acquisition strategies, using the same network architecture and hyperparameters. We further demonstrate its efficiency in sparsely representing transforms along low-dimensional manifolds, resulting in superior immunity to noise and reconstruction artifacts compared with conventional handcrafted reconstruction methods. In addition to improving the reconstruction performance of existing acquisition methodologies, we anticipate accelerating the discovery of new acquisition strategies across modalities as the burden of reconstruction becomes lifted by AUTOMAP and learned-reconstruction approaches.

References
Methods

Knowls

Knowl 1 — AUTOMAP Framework for Domain Transform Reconstruction
model/method
Automated Transform by Manifold Approximation (AUTOMAP) recasts image reconstruction from sensor domain measurements into a data-driven, supervised manifold learning problem. Rather than applying a handcrafted multi-stage inversion pipeline, AUTOMAP learns an end-to-end mapping $f: \mathbb{R}^{n^2} \to \mathbb{R}^{n^2}$ between a sensor domain observation $x \in \mathcal{X}$ and an underlying image domain target $y \in \mathcal{Y}$ , where $\mathcal{X} \subset \mathbb{R}^{n^2}$ and $\mathcal{Y} \subset \mathbb{R}^{n^2}$ are unknown low-dimensional smooth manifolds with $\dim(\mathcal{X}) < n^2$ and $\dim(\mathcal{Y}) < n^2$ .

The mapping operates over the joint manifold $\mathcal{M}_{X,Y} = \mathcal{X} \times \mathcal{Y} = \{(x, f(x)) \in \mathbb{R}^{n^2} \times \mathbb{R}^{n^2} \mid x \in \mathcal{X}, f(x) \in \mathcal{Y}\}$ and is expressed as a composition of coordinate transformations:

$f(x) = \phi_Y \circ g \circ \phi_X^{-1}(x)$

where $\phi_X: \mathcal{Z} \to \mathcal{X}$ and $\phi_Y: \mathcal{Z} \to \mathcal{Y}$ define the local coordinate charts from the intrinsic low-dimensional coordinate space $\mathcal{Z}$ to Euclidean space near $x$ and $y$ , and $g: \mathcal{Z} \to \mathcal{Z}$ is a diffeomorphism representing the between-manifold projection from the sensor domain manifold to the image domain manifold.
Knowl 2 — AUTOMAP Deep Neural Network Architecture
model/method
AUTOMAP implements the manifold domain transformation and sparse feature decompression using a deep feed-forward neural network consisting of fully-connected layers followed by a convolutional autoencoder:
1. Input Vectorization: For an $n \times n$ target image (with $n=128$ ), complex-valued sensor data (such as $k$ -space) is separated into real and imaginary parts and concatenated into a real-valued input vector of size $2n^2 \times 1$ .
2. Fully-Connected Manifold Transformation: The input layer $\text{FC1}$ ( $2n^2$ units) connects to a hidden layer $\text{FC2}$ with $n^2$ units activated by the hyperbolic tangent function ( $\tanh$ ). $\text{FC2}$ connects to a second hidden layer $\text{FC3}$ with $n^2$ units, also activated by $\tanh$ . The output of $\text{FC3}$ is reshaped into a two-dimensional matrix of size $n \times n$ .
3. Convolutional Feature Representation and Deconvolution:
  - Layer $\text{C1}$ : Convolves the $n \times n$ map with 64 filters of size $5 \times 5$ (stride 1) followed by a Rectified Linear Unit (ReLU) nonlinearity.
  - Layer $\text{C2}$ : Convolves the feature maps from $\text{C1}$ with 64 filters of size $5 \times 5$ (stride 1) followed by a ReLU nonlinearity.
  - Output Deconvolutional Layer: Deconvolves the feature maps from $\text{C2}$ using 64 filters of size $7 \times 7$ (stride 1) to synthesize the reconstructed $n \times n$ magnitude (or separated real and imaginary) image.
Knowl 3 — AUTOMAP Training Objective and Optimization
algorithm
The AUTOMAP reconstruction network is trained end-to-end to jointly optimize the domain transformation and the sparse convolutional representation.
Input: Training set of sensor encodings $x_i$ and ground-truth images $y_i$ for $i = 1, \dots, N$
Output: Trained AUTOMAP network parameters $\Theta$
Initialize network weights $\Theta$
for epoch = 1 to 100 do
    for each minibatch $\{(x_b, y_b)\}_{b=1}^{B}$ of size $B = 100$ do
        Apply $1\%$ multiplicative corruption noise to sensor inputs: $\tilde{x}_b = x_b \cdot (1 + \epsilon_b)$ with $\epsilon_b \sim \mathcal{N}(0, 0.01^2)$
        Compute network output: $\hat{y}_b = f_\Theta(\tilde{x}_b)$
        Extract feature map activations $a_{\text{C2}, b}$ from convolutional layer C2
        Compute loss: $\mathcal{L} = \frac{1}{B} \sum_{b=1}^B \| \hat{y}_b - y_b \|_2^2 + \lambda \frac{1}{B} \sum_{b=1}^B \| a_{\text{C2}, b} \|_1$ with $\lambda = 0.0001$
        Update parameters $\Theta$ using RMSProp (learning rate $= 2 \times 10^{-5}$ , momentum $= 0.0$ , decay $= 0.9$ )
    end for
end for
return $\Theta$
The $1\%$ multiplicative input corruption forces the network to learn robust manifold projections rather than relying on explicit additive Gaussian noise models. The $L_1$ penalty on layer C2 feature maps enforces sparsity in the learned convolutional dictionary.
Knowl 4 — Manifold Learning and Joint Denoising-Reconstruction Formulation
theoretical result
Let $X \in \mathcal{X}$ be true sensor data, $\tilde{X}$ be noisy/corrupted sensor observations governed by corruption distribution $P(\tilde{X}|X)$ , and $Y = f(X) \in \mathcal{Y}$ be the true image. The reconstruction model parameterizes the joint distribution of $(Y, X, \tilde{X})$ as:

$Q_{(f,p)}(Y, X, \tilde{X}) = Q(Y \mid X, f) Q(X \mid \tilde{X}, p) P(\tilde{X}) = Q(Y \mid \tilde{X}, (f,p)) P(\tilde{X})$

where $Q(X \mid \tilde{X}, p)$ is a denoising projection in a semiparametric family $\mathbb{Q} = \{ Q(X \mid \tilde{X} = \tilde{x}, p) \mid \mathbb{E}[X] = p(\tilde{X}) \}$ , and $Q(Y \mid X, f)$ is the degenerate reconstruction distribution with point mass at $y = f(x)$ .

Training minimizes the Kullback-Leibler divergence between the empirical distribution $P(Y, X, \tilde{X}) = P(Y \mid X) P(\tilde{X} \mid X) P(X)$ and the model $Q_{(f,p)}(Y, X, \tilde{X})$ :

$\min_{f, p} \mathbb{D}_{\mathrm{KL}}\left(P(Y, X, \tilde{X}) \,\Vert\, Q_{(f,p)}(Y, X, \tilde{X})\right)$

As the KL-divergence tends to zero, the denoising estimator $Q(X \mid \tilde{X}, p) \to P(X \mid \tilde{X})$ and the predictive distribution $Q(Y \mid \tilde{X}, (f,p)) \to P(Y \mid \tilde{X})$ . Because coordinate charts $\phi_X, \phi_Y$ , stochastic projection $p$ , and diffeomorphism $g$ belong to $C^\infty$ , the composite function $\hat{f} = \phi_Y \circ g \circ \phi_X^{-1} \circ p$ is continuously differentiable on a compact subset of $\mathbb{R}^{n^2}$ , which guarantees universal function approximability by a feed-forward neural network.
Knowl 5 — Synthetic Phase Modulation for Complex-Valued Reconstruction
model/method
To train AUTOMAP to reconstruct both image magnitude and image phase from complex-valued sensor data when using public databases (such as the Human Connectome Project) that contain only magnitude images, a synthetic phase modulation method is used:
1. Two-dimensional synthetic phase maps $\theta(u, v)$ are generated from sinusoidal functions with independent spatial frequencies along orthogonal image axes, rotated by a random angle with respect to the coordinate grid.
2. Phase values are scaled and normalized to the range $[0, 2\pi]$ .
3. Each magnitude training image $M(u, v)$ is modulated to produce a complex-valued target image $I(u, v) = M(u, v) \exp(i \theta(u, v))$ .
4. The complex image $I(u, v)$ is transformed via the sensor domain forward model (such as the Fast Fourier Transform) to produce the network input.
5. Separate AUTOMAP networks (or a single network with concatenated outputs) are trained with the resulting inputs to reconstruct the target real/magnitude and imaginary/phase components, enabling generalization to in vivo complex-valued acquisitions without requiring real phase training datasets.
Knowl 6 — AUTOMAP Reconstruction Performance Across Multiple Encoding Strategies
empirical result
Using an identical network architecture and hyperparameter set, AUTOMAP was evaluated against standard domain-specific reconstruction algorithms across four distinct sensor encoding modalities:
1. Discrete Radon Transform (180 angles, 185 parallel rays, with $40\text{ dB SNR}$ additive white Gaussian noise): Compared against the Kaczmarz-iterative Algebraic Reconstruction Technique (ART, 10 iterations), AUTOMAP suppressed white-noise amplification artifacts present in ART.
2. Spiral Non-Cartesian $k$ -space (10-interleave trajectory, variable density factor $\alpha=1$ , undersampling $R=1/1.2$ , with $25\text{ dB SNR}$ additive noise): Compared against single-coil Conjugate-Gradient SENSE with NUFFT regridding (30 iterations), AUTOMAP eliminated regridding-induced noise ringing.
3. Poisson-Disc Undersampled Cartesian $k$ -space ( $40\%$ undersampling, with $30\text{ dB SNR}$ additive noise): Compared against compressed sensing with a wavelet sparsifying transform (BART implementation, $\lambda=0.01$ ), AUTOMAP avoided complex noise-aliasing artifacts.
4. Misaligned Cartesian $k$ -space (random shifts up to $\pm 3$ samples per readout line): Compared against the standard 2D Inverse Fast Fourier Transform (2D-IFFT), AUTOMAP reconstructed artifact-free images, eliminating severe trajectory misalignment ghosting and aliasing.
Knowl 7 — Spontaneous Sparsification and Spatial Weight Organization in Hidden Representations
empirical result
Analysis of the hidden layers and learned weights of AUTOMAP for Cartesian $k$ -space reconstruction demonstrates two emergent properties:
1. Emergent Activation Sparsity: When reconstructing the Cartesian $k$ -space of a brain image, hidden-layer $\text{FC2}$ activations become progressively sparser as the training corpus shifts from random Gaussian noise to generic natural images (ImageNet), and sparsest when trained on domain-specific brain images (Human Connectome Project). This activation sparsity emerges naturally without an explicit $L_1$ penalty on the fully-connected layers.
2. Spatial Autocorrelation in Weight Space: Visualizing the $n^2$ $n^{2}$ -dimensional weight vectors connecting $\text{FC2}$ $FC2$ to each pixel in $\text{FC3}$ $FC3$ via 3D t-SNE reveals that:
  - Weights from a noise-trained network show a spatially disordered, uniform distribution (corresponding to a pure, context-agnostic Fourier Transform).
  - Weights from a generic natural-image trained network exhibit local spatial clustering matching pixel adjacency.
  - Weights from a domain-specific brain-image trained network organize into a continuous, coherent two-dimensional manifold/sheet in 3D embedding space, reflecting the strong spatial autocorrelation of anatomical structures.
Knowl 8 — Multi-Task Sensor Domain Encoding and Dataset Configurations
experimental setup
The experimental setups and data preprocessing used to train and evaluate AUTOMAP across forward models were:
- Generic Natural Images: 10,000 images from ImageNet ('Animal', 'Plant', 'Scene') cropped to central $256 \times 256$ , subsampled to $128 \times 128$ , converted to grayscale Y-channel luminance, augmented by $90^\circ$ rotations, zero-mean centered, and normalized by dataset maximum.
- Brain MRI Images: 50,000 T1-weighted 3D MPRAGE slices from 131 subjects from the MGH-USC Human Connectome Project (HCP; $\text{TR}=2530\text{ ms}$ , $\text{TE}=1.15\text{ ms}$ , $\text{TI}=1100\text{ ms}$ , $\text{FA}=7.0^\circ$ , $\text{BW}=651\text{ Hz/Px}$ on Siemens 3T Skyra). Symmetrically tiled to $256 \times 256$ with four reflections and randomly cropped to $128 \times 128$ for translation invariance.
- Sensor Encodings:
  - Radon: Discrete Radon Transform with 180 projection angles and 185 parallel rays.
  - Spiral $k$ -space: NUFFT with 10-interleave spiral trajectory, variable density $\alpha=1$ , undersampling $R=1/1.2$ .
  - Poisson-disc Cartesian: $40\%$ undersampling generated with the Berkeley Advanced Reconstruction Toolbox (BART).
  - Misaligned Cartesian: FFT with each readout line randomly displaced along the readout direction by up to $\pm 3$ samples.
  - In Vivo Test Data: Spin-echo sequence on Siemens 3T Trio ( $\text{TR}=3110\text{ ms}$ , $\text{TE}=23.0\text{ ms}$ , matrix $208 \times 256$ , slice thickness $3\text{ mm}$ , 12-channel head coil SVD coil-compressed to central $128 \times 128$ ).

Coverage note — None was omitted; all primary theoretical formulations, network architectural details, optimization procedures, experimental configurations, and empirical findings are fully covered.

References

1.Arridge, S. R. & Schweiger, M. Image reconstruction in optical tomography. Philosophical Transactions of the Royal Society B: Biological Sciences 352, 717–726 (1997).
2.Jakowatz, C. V. J., Wahl, D. E., Eichel, P. H., Ghiglia, D. C. & Thompson, P. A. Spotlight-Mode Synthetic Aperture Radar: A Signal Processing Approach. (1996).
3.Fessler, J. A. & Sutton, B. P. Nonuniform fast fourier transforms using min-max interpolation. IEEE Trans. Signal Process. 51, 560–574 (2003).
4.Kak, A. C. & Slaney, M. Principles of Computerized Tomographic Imaging. (Society for Industrial and Applied Mathematics, 2012). doi:10.1137/1.9780898719277
5.Kinahan, P. E., Defrise, M. & Clackdoyle, R. in Emission Tomography 421–442 (Elsevier, 2004). doi:10.1016/b978-012744482-6.50023-5
6.Szabo, T. L. Diagnostic Ultrasound Imaging: inside out. Academic Press (2004).
7.Gull, S. F. & Daniell, G. J. Image reconstruction from incomplete and noisy data. Nature 272, 686–690 (1978).
8.Hinton, G. et al. Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups. IEEE Signal Process. Mag. 29, 82–97 (2012).
9.Krizhevsky, A., Sutskever, I. & Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. Advances in neural … 1097–1105 (2012).
10.Gilbert, C. D., Sigman, M. & Crist, R. E. The Neural Basis of Perceptual Learning. Neuron 31, 681–697 (2001).
11.Lu, Z.-L., Hua, T., Huang, C.-B., Zhou, Y. & Dosher, B. A. Visual perceptual learning. Neurobiology of Learning and Memory 95, 145–151 (2011).
12.Lustig, M. & Pauly, J. M. SPIRiT: Iterative self-consistent parallel imaging reconstruction from arbitrary k-space. Magn. Reson. Med. 64, 457-471 (2010).
13.Pruessmann, K. P., Weiger, M., Börnert, P. & Boesiger, P. Advances in sensitivity encoding with arbitrary k‐space trajectories. Magn. Reson. Med. 46, 638–651 (2001).
14.Vincent, P., Larochelle, H., Bengio, Y. & Manzagol, P.-A. Extracting and composing robust features with denoising autoencoders. Proceedings of the 25th … 1096–1103 (ACM, 2008). doi:10.1145/1390156.1390294
15.Beister, M., Kolditz, D. & Kalender, W. A. Iterative reconstruction methods in X-ray CT. Physica Medica 28, 94–108 (2012).
16.Lustig, M., Donoho, D. & Pauly, J. M. Sparse MRI: The application of compressed sensing for rapid MR imaging. Magn. Reson. Med. 58, 1182–1195 (2007).
17.Fan, Q. et al. MGH–USC Human Connectome Project datasets with ultra-high b-value diffusion MRI. NeuroImage 124, 1108–1114 (2016).
18.Jia Deng et al. ImageNet: A large-scale hierarchical image database. in 248–255 (IEEE, 2009). doi:10.1109/CVPRW.2009.5206848
19.Hornik, K., Stinchcombe, M. & White, H. Multilayer feedforward networks are universal approximators. Neural Networks 2, 359–366 (1989).
20.Poggio, T. & Girosi, F. Networks for approximation and learning. Proc. IEEE 78, 1481–1497 (1990).
21.Di Carli, M. F. & Lipton, M. J. Cardiac PET and PET/CT Imaging. (Springer New York, 2007). doi:10.1007/978-0-387-38295-1
22.Yang, Z. & Jacob, M. Mean square optimal NUFFT approximation for efficient non-Cartesian MRI reconstruction. Journal of Magnetic Resonance 242, 126–135 (2014).
23.Virtue, P. & Lustig, M. On the Empirical Effect of Gaussian Noise in Undersampled MRI Reconstruction. arXiv preprint arXiv:1610.00410 (2016).
24.Brown, R. W., Cheng, Y. C. N., Haacke, E. M., Thompson, M. R. & Venkatesan, R. Magnetic Resonance Imaging. (2014).
25.Bennett, P. J., Gold, J. & Sekuler, A. B. Signal but not noise changes with perceptual learning : Abstract : Nature. Nature 402, 176–178 (1999).
26.Wright, J. et al. Sparse Representation for Computer Vision and Pattern Recognition. Proc. IEEE 98, 1031–1044 (2010).
27.Maaten, L. V. D. & Hinton, G. Visualizing Data using t-SNE. Journal of Machine Learning Research 9, 2579–2605 (2008).
28.Getis, A. Spatial Autocorrelation, in Handbook of Applied Spatial Analysis 255–278 (Springer Berlin Heidelberg, 2010). doi:10.1007/978-3-642-03647-7_14
29.Kubo, T. et al. Radiation Dose Reduction in Chest CT: A Review. American Journal of Roentgenology 190, 335–343 (2008).
30.Daigle, O., Djazovski, O., Laurin, D., Doyon, R. & Artigau, É. Characterization results of EMCCDs for extreme low-light imaging. SPIE Astronomical Telescopes+ Instrumentation 8453, 845303–845303–9 (2012).
31.Girard, J. N. et al. Sparse representations and convex optimization as tools for LOFAR radio interferometric imaging. J. Inst. 10, C08013 (2015).
32.Lebed, E., Sarunic, M. V., Beg, M. F. & Mackenzie, P. J. Rapid volumetric OCT image acquisition using Compressive Sampling. Opt. Express, OE 18, 21003–21012 (2010).
33.Kim, D. H., Adalsteinsson, E. & Spielman, D. M. Simple analytic variable density spiral design. Magn. Reson. Med. 50, 214–219 (2003).
34.Uecker, M. et al. Berkeley advanced reconstruction toolbox. in Proceedings of the International Society for Magnetic Resonance in Medicine, Toronto (2015).
35.Abadi, M. et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems. arXiv preprint arXiv:1603.04467 (2016).
36.Nair, V. & Hinton, G. E. Rectified linear units improve restricted boltzmann machines. in (2010).
37.Makhzani, A. & Frey, B. J. Winner-Take-All Autoencoders. Advances in Neural Information Processing Processing Systems. 2791–2799 (2015).
38.Hansen, P. C. & Saxild-Hansen, M. AIR Tools — A MATLAB package of algebraic iterative reconstruction methods. Journal of Computational and Applied Mathematics 236, 2167–2178 (2012).
39.Bengio, Y., Courville, A. & Vincent, P. Representation Learning: A Review and New Perspectives. IEEE Trans. Pattern Anal. Mach. Intell. 35, 1798–1828 (2013).
40.Yu, K. & Zhang, T. Improved Local Coordinate Coding using Local Tangents. in 1215–1222 (2010).
41.Anderes, E. & Coram, M. A general spline representation for nonparametric and semiparametric density estimates using diffeomorphisms. arXiv:1205.5314 [stat] (2012).
42.Zhang, M., Singh, N. & Fletcher, P. T. Bayesian Estimation of Regularization and Atlas Building in Diffeomorphic Image Registration. in 37–48 (Springer, Berlin, Heidelberg, 2013). doi:10.1007/978-3-642-38868-2_4
43.Fishbaugh, J., Prastawa, M., Gerig, G. & Durrleman, S. Geodesic image regression with a sparse parameterization of diffeomorphisms. Geometric science of information : first international conference, GSI 2013, Paris, France, August 28-30, 2013 : proceedings / Frank Nielsen, Frederic Barbaresco (eds.). Geometric Science of Information (1st : 2013 : Paris, France) 8085, 95–102 (2013).
44.Bernstein, A., Kuleshov, A. & Yanovich, Y. Manifold Learning in Regression Tasks. in 414–423 (Springer, Cham, 2015). doi:10.1007/978-3-319-17091-6_36
45.Hornik, K. Approximation capabilities of multilayer feedforward networks. Neural Networks 4, 251–257 (1991).

Citation

MLA

Zhu, B., et al. “Image Reconstruction by Domain-transform Manifold Learning”. Nature, vol. 555, no. 7697, 2018, pp. 487–92, https://doi.org/10.1038/nature25988.

APA

Zhu, B., Liu, J. Z., Cauley, S. F., Rosen, B. R., & Rosen, M. S. (2018). Image reconstruction by domain-transform manifold learning. Nature, 555(7697), 487–492. https://doi.org/10.1038/nature25988

Chicago

Zhu, B., J. Z. Liu, S. F. Cauley, B. R. Rosen, and M. S. Rosen. 2018. “Image Reconstruction by Domain-transform Manifold Learning”. Nature 555 (7697): 487–92. https://doi.org/10.1038/nature25988.

Harvard

Zhu, B. et al. (2018) “Image reconstruction by domain-transform manifold learning”, Nature, 555(7697), pp. 487–492. Available at: https://doi.org/10.1038/nature25988.

Vancouver

1. Zhu B, Liu JZ, Cauley SF, Rosen BR, Rosen MS (2018) Image reconstruction by domain-transform manifold learning. Nature 555:487–492

BibTeX

@article{Zhu_2018, title={Image reconstruction by domain-transform manifold learning}, volume={555}, ISSN={1476-4687}, url={http://dx.doi.org/10.1038/nature25988}, DOI={10.1038/nature25988}, number={7697}, journal={Nature}, publisher={Springer Science and Business Media LLC}, author={Zhu, Bo and Liu, Jeremiah Z. and Cauley, Stephen F. and Rosen, Bruce R. and Rosen, Matthew S.}, year={2018}, month=Mar, pages={487–492} }

Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Image reconstruction by domain-transform manifold learning

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — AUTOMAP Framework for Domain Transform Reconstruction

Knowl 2 — AUTOMAP Deep Neural Network Architecture

Knowl 3 — AUTOMAP Training Objective and Optimization

Knowl 4 — Manifold Learning and Joint Denoising-Reconstruction Formulation

Knowl 5 — Synthetic Phase Modulation for Complex-Valued Reconstruction

Knowl 6 — AUTOMAP Reconstruction Performance Across Multiple Encoding Strategies

Knowl 7 — Spontaneous Sparsification and Spatial Weight Organization in Hidden Representations

Knowl 8 — Multi-Task Sensor Domain Encoding and Dataset Configurations

References

Citation

BibTeX

Source Code

Access the Paper