Compressed Sensing

The core requirement for compressed sensing is that the signal must be sparse or compressible in some known basis. This means the signal can be represented with only a small number of non-zero coefficients when transformed.

• Common transform domains: Wavelets, Fourier basis, discrete cosine transform, or learned dictionaries
• Medical imaging example: MR images are sparse in the wavelet domain; angiograms are sparse in the spatial gradient domain
• A signal with only 5% non-zero wavelet coefficients can be reconstructed from dramatically under-sampled k-space data
• The sparsity level (k) directly determines the minimum number of measurements required for successful recovery

Compressed sensing requires that the measurement basis and the sparsity basis be incoherent—meaning they are maximally uncorrelated. This ensures each measurement captures global information rather than local redundancy.

• Incoherence property: Low-coherence sampling spreads aliasing artifacts into noise-like patterns rather than coherent structures
• Random under-sampling of k-space in MRI is a practical implementation—it creates incoherent artifacts that can be removed by non-linear reconstruction
• Variable-density sampling: More samples are acquired near the k-space center (low frequencies) with decreasing density toward the periphery
• The restricted isometry property (RIP) provides the mathematical guarantee that a sensing matrix preserves signal geometry for sparse vectors

Recovery from under-sampled data requires solving a non-linear optimization problem that enforces both sparsity and data consistency. This is fundamentally different from linear methods like Filtered Back Projection.

• L1-norm minimization: Minimizing the L1 norm of the transform coefficients promotes sparsity, unlike L2 (Tikhonov) regularization which produces overly smooth solutions
• Constrained optimization form: min ||Ψx||₁ subject to ||Ax - y||₂ < ε where Ψ is the sparsifying transform, A is the sensing matrix, and y is the acquired data
• Iterative soft-thresholding algorithms progressively denoise the estimate while maintaining fidelity to measured data
• Total Variation (TV) regularization is often combined with wavelet sparsity to suppress noise while preserving edges in medical images

The acceleration factor (R) quantifies how much the acquisition is under-sampled relative to fully-sampled Nyquist-rate data. Higher R values mean faster scans but more challenging reconstructions.

• Typical acceleration factors: 2x–4x for routine clinical MRI, 8x–16x for advanced research protocols
• Cartesian under-sampling: Regularly spaced skipped phase-encoding lines; simple but produces coherent aliasing
• Radial and spiral trajectories: Inherently incoherent sampling patterns that are more robust for compressed sensing
• Poisson-disk sampling: A variable-density random pattern that enforces a minimum distance between samples, optimizing incoherence while respecting hardware gradient constraints

Modern implementations combine classical compressed sensing theory with deep neural networks to push acceleration factors beyond traditional limits while preserving diagnostic quality.

• Unrolled optimization networks: Each iteration of an iterative reconstruction algorithm is mapped to a neural network layer, with learnable parameters replacing hand-tuned regularization
• GAN-based reconstruction: Generative adversarial networks learn to produce realistic high-resolution images from under-sampled inputs, effectively learning the image prior from data
• End-to-end mapping: Networks like AUTOMAP learn the entire reconstruction pipeline directly from k-space to image domain
• Clinical deployment: FDA-cleared DLR systems (e.g., GE AIR Recon DL, Siemens Deep Resolve) now routinely achieve 4x–8x acceleration in musculoskeletal and neuroimaging protocols

A critical insight of compressed sensing is that random under-sampling transforms coherent aliasing into incoherent, noise-like artifacts that can be separated from the true signal during reconstruction.

• Coherent aliasing: Regular under-sampling produces discrete, structured ghosting artifacts (e.g., wraparound in phase-encoding direction)
• Incoherent aliasing: Random under-sampling spreads artifact energy diffusely across the image, resembling additive noise
• Iterative thresholding exploits this property by shrinking small coefficients (noise) while preserving large coefficients (signal) in the transform domain
• This principle is why variable-density random sampling is essential—it ensures the point spread function of the under-sampling pattern has minimal coherent side lobes

The fundamental requirement for Compressed Sensing is sparsity—the signal must have a concise representation in a known transform domain.

• Wavelet Transform: Excellent for piecewise-smooth signals like medical images.
• Total Variation (TV): Minimizes the sum of gradient magnitudes, preserving sharp edges.
• Discrete Cosine Transform (DCT): Basis for JPEG; useful for highly correlated data.
• Dictionary Learning: Adaptively learns sparse representations from data itself.

Without sparsity, accurate reconstruction from under-sampled data is mathematically impossible.

To avoid aliasing artifacts, under-sampling must be incoherent with the sparsity basis. This means sampling patterns must appear noise-like in the sparsifying domain.

• Random Cartesian Under-sampling: Pseudo-randomly skipping phase-encoding lines in MRI.
• Radial/Spiral Trajectories: Naturally incoherent with standard image bases.
• Variable Density Sampling: Denser sampling at the k-space center (high energy), sparser at the periphery.

Incoherence ensures that aliasing artifacts spread like noise rather than coherent ghosts.

Recovery from under-sampled data requires solving a convex optimization problem, typically using L1-norm minimization.

• L1-Minimization: The convex surrogate for the NP-hard L0 problem; promotes sparsity.
• Iterative Soft-Thresholding (ISTA): Proximal gradient method alternating between data consistency and soft-thresholding.
• ADMM (Alternating Direction Method of Multipliers): Splits the problem into simpler sub-problems for faster convergence.
• Compressed Sensing MRI (CS-MRI): Formulation: min ||Ψx||₁ s.t. ||Fᵤx - y||₂ < ε.

The Restricted Isometry Property (RIP) provides a sufficient condition guaranteeing stable recovery. A sensing matrix A satisfies RIP of order k if it nearly preserves the L2-norm of all k-sparse vectors.

• RIP Constant (δₖ): Measures the maximum deviation from perfect isometry.
• Theoretical Guarantee: If δ₂ₖ < √2 - 1, L1-minimization recovers the exact sparse signal.
• Gaussian Random Matrices: Universally satisfy RIP with high probability.
• Practical Implication: Validates that random under-sampling enables robust reconstruction.

Modern approaches replace hand-crafted sparsity priors with data-driven priors learned by deep neural networks, often outperforming classical CS.

• Unrolled Networks (e.g., ADMM-Net): Map iterative optimization steps to neural network layers.
• Variational Networks: Learn the regularization term and gradient steps from data.
• GAN-based Reconstruction: Use adversarial loss to enforce perceptual realism.
• End-to-End Mapping: Directly learn the mapping from under-sampled k-space to the artifact-free image.

DLR achieves higher acceleration factors (R=8-10x) with superior perceptual quality.

The acceleration factor (R) quantifies the reduction in acquired data relative to Nyquist-rate sampling. An R=4 scan is four times faster.

• SSIM (Structural Similarity Index): Measures perceptual similarity to the fully-sampled reference.
• PSNR (Peak Signal-to-Noise Ratio): Quantifies reconstruction error in decibels.
• NMSE (Normalized Mean Squared Error): Scale-invariant error metric.
• Clinical Equivalence: The ultimate metric—do radiologists make identical diagnostic decisions?

High acceleration factors must not compromise diagnostic accuracy.