Loading field map
Medical imaging is the set of techniques used to create visual representations of the interior of a living body for clinical diagnosis, treatment planning, and biomedical research. It is a subfield of biomedical engineering that sits at the intersection of physics, signal processing, computer science, and physiology. The central problem of medical imaging is not simply "taking pictures" but rather inferring the physical state of tissues and organs from indirect measurements, often without disturbing the body. Every imaging modality is a solution to this problem, and each solution involves a different trade-off among spatial resolution, contrast, speed, cost, and safety.
All medical images are maps of some physical property of tissue. X-ray imaging maps the attenuation of X-ray photons; magnetic resonance imaging (MRI) maps the density and magnetic relaxation properties of hydrogen nuclei; ultrasound maps the reflection of sound waves at boundaries between tissues with different acoustic impedances; nuclear medicine maps the distribution of a radioactive tracer that is metabolically taken up by specific cells. The image is never a direct photograph of anatomy; it is a reconstruction of a physical parameter from measurements made outside the body.
Two quantities govern the usefulness of any image. Spatial resolution is the smallest distance at which two adjacent structures can be distinguished as separate. Contrast is the difference in signal between two tissues of interest relative to background noise. These two are in constant tension. Improving resolution often requires more radiation dose, longer acquisition times, or stronger magnetic fields, all of which have practical limits. Improving contrast often requires injecting contrast agents or using tracers, which carry their own risks. The history of medical imaging is largely a history of clever ways to push both quantities within acceptable safety and cost constraints.
The field began in 1895 with Wilhelm Röntgen's discovery of X-rays, which immediately found clinical use for visualizing bones and, later, with the aid of contrast agents, blood vessels and the gastrointestinal tract. For the first half of the twentieth century, projection radiography—a single X-ray image of the body from one angle—was the only imaging method. Its limitation is fundamental: a two-dimensional projection collapses three-dimensional anatomy, so overlapping structures obscure one another.
The first true revolution came in the 1970s with computed tomography (CT). CT uses an X-ray source and detector that rotate around the patient, acquiring hundreds of projection images from different angles. A computer then solves the mathematical problem of reconstructing a three-dimensional volume from these projections. The underlying mathematics, known as tomographic reconstruction, had been developed decades earlier in radio astronomy and electron microscopy, but it required the computing power of the 1970s to become clinically practical. CT provided, for the first time, cross-sectional images of the body with excellent spatial resolution and clear contrast between bone, soft tissue, and air.
Magnetic resonance imaging emerged in the late 1970s and early 1980s from a different physical principle. Certain atomic nuclei, particularly hydrogen, behave like tiny magnets. When placed in a strong magnetic field and excited by radio-frequency pulses, they emit signals whose frequency depends on the local magnetic field. By applying magnetic field gradients, the signal can be spatially encoded, and a computer reconstructs an image. MRI's great advantage is its extraordinary soft-tissue contrast—it can distinguish gray matter from white matter in the brain, or cartilage from fluid in a joint—without any ionizing radiation. Its disadvantages are cost, long acquisition times, and the need for very strong magnets.
Ultrasound developed in parallel, beginning in the 1950s. It uses high-frequency sound waves that are emitted into the body and reflect at boundaries between tissues with different acoustic properties. The time delay of the reflected echo gives the depth of the reflecting structure. Ultrasound is real-time, inexpensive, portable, and uses no ionizing radiation, but its image quality is operator-dependent and it cannot image through bone or air-filled structures like the lungs.
Nuclear medicine, including single-photon emission computed tomography (SPECT) and positron emission tomography (PET), takes a different approach. Instead of imaging anatomy, it images physiology. A radioactive compound is injected, and it accumulates in tissues according to metabolic activity. The emitted radiation is detected outside the body, and a map of tracer concentration is reconstructed. PET and SPECT have poor spatial resolution compared to CT or MRI, but they provide unique functional information—for example, which parts of the brain are active during a task, or where cancer cells are metabolically active.
The field is best organized around a fundamental distinction between two paradigms that have coexisted since the 1970s and continue to define the field's structure.
Anatomical imaging aims to depict the structure of the body—the shape, size, and position of organs, bones, and lesions. CT and MRI are the primary anatomical modalities, with ultrasound playing a major role in specific applications. The central question is: Where is the abnormality, and what does it look like? The methods are designed to maximize spatial resolution and structural contrast. CT excels at bone and lung imaging; MRI excels at soft tissues; ultrasound excels at real-time assessment of moving structures like the heart or fetus.
Functional imaging aims to depict the physiological activity of tissues—blood flow, metabolism, receptor binding, or neural activation. PET, SPECT, and functional MRI (fMRI) are the primary functional modalities. The central question is: How active is this tissue, and is that activity abnormal? The methods are designed to detect changes in tracer uptake or blood oxygenation that correlate with cellular activity. Functional imaging has much poorer spatial resolution than anatomical imaging, but it provides information that anatomy alone cannot: a tumor may be anatomically identical to surrounding tissue but metabolically hyperactive; a brain region may look normal on CT but show reduced blood flow in a stroke.
These two paradigms are not rivals but complements. Modern clinical practice almost always combines them. PET/CT scanners, which acquire both images in a single session, are now standard in oncology because the CT provides the anatomical context for the PET's metabolic signal. Similarly, MRI can be used to acquire both anatomical and functional images in the same session, allowing a surgeon to see both where a brain tumor is and which nearby regions are responsible for speech or movement.
A second organizing axis of the field is the mathematical problem of image reconstruction. Every tomographic modality—CT, MRI, PET, SPECT—faces the same abstract challenge: the measured data are not images but projections or frequency samples of the object, and the image must be computed from these indirect measurements.
The earliest and most widely used solution is filtered backprojection. In CT, each projection is a one-dimensional set of line integrals through the body. The reconstruction algorithm smears each projection back across the image plane and sums the smears, then applies a filter to correct for the blurring that this smearing introduces. Filtered backprojection is fast and robust, and it remains the workhorse of clinical CT. Its limitation is that it amplifies noise and produces streak artifacts when the number of projections is limited.
The alternative is iterative reconstruction. Instead of computing the image in one pass, iterative methods start with an initial guess, simulate what measurements that guess would produce, compare the simulation to the actual measurements, and update the guess to reduce the difference. This process repeats until convergence. Iterative methods can incorporate detailed models of the physics of the imaging system—including noise statistics, scatter, and detector response—and they can produce better images at lower radiation doses. Their disadvantage is computational cost, which has only recently become manageable with modern graphics processing units. Iterative reconstruction is now standard in PET and is increasingly used in CT to reduce radiation dose.
A third approach, which has grown rapidly since the 2010s, is deep learning–based reconstruction. Neural networks are trained on large datasets of paired low-quality and high-quality images, or on the physics of the imaging system, to map raw measurements directly to images. These methods can produce images that appear to have both high resolution and low noise, but they raise a fundamental concern: the network may hallucinate structures that are not actually present, or suppress real structures that are rare in the training data. The field is actively debating how to validate these methods and whether they can be trusted for clinical decisions.
The contemporary field is defined by three trends that cut across the anatomical/functional distinction.
Multimodal fusion is the integration of images from different modalities into a single coherent representation. This can be done at the hardware level, as in PET/CT and PET/MRI scanners, or at the software level, by aligning images acquired separately through a process called registration. Fusion is valuable because no single modality provides all the needed information. A PET/CT scan shows both the metabolic activity of a tumor and its exact anatomical location; a fused MRI and CT scan of the head shows both soft-tissue detail and bone anatomy for surgical planning.
Quantitative imaging is the movement from visual interpretation to numerical measurement. Instead of a radiologist looking at an image and judging whether a lesion is present, quantitative imaging extracts numbers: the apparent diffusion coefficient of water in a tumor, the standardized uptake value of a tracer in a tissue, the fractional anisotropy of white matter tracts in the brain. These numbers can be compared across patients, tracked over time, and correlated with outcomes. The challenge is standardization—ensuring that the same measurement on different scanners and at different institutions produces the same number.
Artificial intelligence is transforming both image interpretation and image acquisition. Machine learning algorithms can detect suspicious lesions on chest X-rays, segment organs on CT scans, estimate the age of a brain from an MRI, and predict the likelihood of a tumor being malignant from its imaging features. These tools are not replacing radiologists but are increasingly used as decision support, triage tools, and quality control. The central unresolved questions are regulatory: how to validate an algorithm's performance across diverse patient populations, how to ensure it does not encode biases, and how to assign responsibility when it makes an error.
The field's enduring challenge remains what it has always been: extracting reliable information about the living body from indirect physical measurements. Each new modality, each improvement in reconstruction, and each application of machine learning is a new answer to that same question. The practical stakes are enormous—imaging guides nearly every major medical decision, from cancer staging to stroke treatment to prenatal care—and the field continues to evolve as physics, mathematics, and computer science provide new tools.