Magnetic Resonance Imaging

Magnetic resonance imaging (MRI) is a medical imaging modality that constructs images of the body's interior not by passing radiation through the patient, as in X-ray or CT, but by manipulating the quantum-mechanical spin states of hydrogen nuclei using magnetic fields and radio waves. The technique is a triumph of applied physics and signal processing: it translates the nuclear magnetic resonance phenomenon — a quantum effect discovered in the 1930s — into clinically useful images through a sophisticated encoding and reconstruction pipeline that is, at its core, a spatial sampling problem in the Fourier domain.

The physical basis of MRI is the Zeeman splitting of hydrogen nuclear spin states in a strong static magnetic field. When a radiofrequency pulse at the Larmor frequency tips these spins out of alignment, the precessing magnetization induces a voltage in a receiver coil. The amplitude of this voltage encodes the density of hydrogen protons at each location. The challenge is spatial localization: unlike a camera, which captures light from different directions simultaneously, an RF coil receives the sum of signals from every excited proton in the body. The solution is frequency encoding.

K-Space and Fourier Encoding

By applying linear magnetic field gradients, the MRI scanner imposes a spatially varying Larmor frequency. Protons at different positions precess at different rates, and their combined signal is the Fourier transform of the spin density distribution. The received signal is not an image but a trajectory through k-space — the spatial frequency domain whose coordinates are determined by the strength and duration of the applied gradients. To reconstruct an image, one must sample k-space sufficiently and compute the inverse Fourier transform.

This structure makes MRI structurally identical to Aperture Synthesis in radio astronomy. In both cases, one cannot directly measure the image domain; instead, one samples the Fourier domain and reconstructs. In radio astronomy, the baselines between antennas determine the k-space samples; in MRI, the gradient waveforms do. Both face the same information-theoretic constraints: the Nyquist-Shannon sampling theorem sets the minimum sampling density, and non-uniform or sparse sampling introduces artifacts unless carefully regularized. The compressed sensing revolution in MRI — enabling scan acceleration by sampling k-space far below the Nyquist rate — is precisely the same mathematical framework that enables sparse recovery in radio interferometry and single-pixel cameras.

From Signal to Image: The Reconstruction Pipeline

The raw data from an MRI scanner are not pixels but frequency-domain samples. The reconstruction pipeline applies inverse Fourier transforms, density compensation for non-Cartesian trajectories, and often iterative algorithms that incorporate prior knowledge about image structure. Modern scanners use parallel imaging — arrays of receiver coils with spatially varying sensitivities — to undersample k-space and recover missing data through the additional spatial encoding provided by the coil profiles. This is a distributed sensing architecture: each coil is an independent sensor, and the final image is a joint estimation problem across all sensors.

The pipeline exemplifies the modern paradigm of computational imaging, where the instrument is inseparable from the algorithm. The image does not exist until the computation is complete. This blurs the boundary between hardware and software, between physics and signal processing, between acquisition and inference. An MRI scanner is not a camera; it is a physics experiment coupled to an inverse problem solver.

Information Limits and Clinical Trade-offs

Every MRI scan is a negotiation between three scarce resources: time, signal-to-noise ratio, and resolution. Longer scans collect more signal and finer k-space samples but are expensive, uncomfortable, and subject to motion artifacts. Higher field strengths increase signal but also shift resonant frequencies in ways that complicate reconstruction. Faster scans sacrifice sampling density and rely on ever more aggressive regularization.

These trade-offs are not merely engineering constraints. They are information-theoretic necessities. The mutual information between the underlying anatomy and the acquired data is bounded by the sampling pattern, the noise level, and the encoding capacity of the coil array. No reconstruction algorithm — however sophisticated — can extract information that was never encoded. The clinical promise of deep learning in MRI reconstruction must be understood within these limits: a neural network can interpolate, denoise, and regularize, but it cannot violate the data processing inequality.

MRI is often presented as a physics technology, but this framing conceals its true nature. Magnetic resonance imaging is a distributed Fourier sampling system whose physical front-end happens to manipulate nuclear spins. The same reconstruction mathematics applies to radio interferometry, synthetic aperture radar, and computed tomography. The field that understands this unity — that sees MRI, the Event Horizon Telescope, and CT as variants of the same inverse problem — will advance faster than the field that treats each as a separate specialty. The specialization of medical imaging from signal processing has been a historical accident driven by clinical funding structures, not by intellectual necessity. The aperture is dead in radiology too.

K-Space and Fourier Encoding

From Signal to Image: The Reconstruction Pipeline

Information Limits and Clinical Trade-offs

See Also