Loading field map
Imaging physics is the subfield that asks a deceptively simple question: what physical interaction between the body and an energy source can be turned into a diagnostic image? The answer has never been obvious, and each major framework in the history of the subfield represents a different bet on which physical principle—absorption, emission, reflection, resonance, or statistical inference—should be the foundation of an image. These bets have shaped not only the machines we build but also what we think an image is: a shadow, a map of radioactivity, a slice of attenuation coefficients, a relaxation time map, or a learned representation from data.
The first framework, Projection Radiography, emerged directly from Röntgen's discovery of X-rays in 1895. Its physical model was elegantly simple: X-rays travel in straight lines through the body, and different tissues absorb them to different degrees. The image was a two-dimensional projection—a shadow—of three-dimensional anatomy. For over seventy years, this framework dominated medical imaging. Its core commitment was that a single, static projection, recorded on film, contained sufficient diagnostic information. The physics was absorption-based, the geometry was parallel-beam or fan-beam projection, and the image formation was purely analog. This framework did not need reconstruction; the image was the direct recording of transmitted intensity. Its limitations—superimposition of structures, poor soft-tissue contrast—were accepted as inherent to the method.
Beginning in the 1950s, two frameworks introduced entirely different physical principles, breaking the monopoly of X-ray absorption.
Nuclear Medicine Imaging replaced the external X-ray source with an internal one. A radioactive tracer was injected into the patient, and the image was formed by detecting gamma rays emitted from within the body. The physics shifted from transmission to emission. The image was no longer a shadow but a functional map of tracer distribution, revealing physiology rather than anatomy. This framework coexisted with Projection Radiography from the start; it did not replace it but carved out a new domain—molecular and metabolic imaging—that absorption-based methods could not reach.
Ultrasound Imaging introduced yet another physical interaction: the reflection of high-frequency sound waves at tissue boundaries. Instead of ionizing radiation, it used mechanical waves. The image was formed by measuring the time delay and amplitude of returning echoes. This framework was fundamentally different from both X-ray and nuclear methods: it was real-time, non-ionizing, and sensitive to tissue interfaces rather than density or tracer concentration. Ultrasound Imaging coexisted with the other two frameworks, offering a complementary view of anatomy and blood flow. Its physics of acoustic impedance mismatch and Doppler shift gave it unique strengths in obstetrics, cardiology, and vascular imaging.
The 1970s brought a conceptual earthquake. Computed Tomography (CT) abandoned the projection entirely as the final image. Instead, it treated the image as a mathematical reconstruction from multiple projections taken at different angles. The physical model remained X-ray absorption, but the framework for forming the image was now computational. The body was represented as a grid of attenuation coefficients, and the task was to solve an inverse problem: given the measured projections, estimate the grid.
The first dominant methodological school within CT was Filtered Back Projection (FBP) . FBP was an analytical, closed-form solution to the reconstruction problem. It assumed that the projection data were complete, noise-free, and that the physics of X-ray attenuation was linear. Its great strength was speed: FBP could reconstruct an image in seconds, making clinical CT practical. Its weakness was that it treated noise and physical imperfections (beam hardening, scatter) as nuisances to be ignored or corrected before reconstruction. FBP became the default reconstruction framework for CT and for early PET and SPECT, lasting as the clinical standard for about thirty years.
Also emerging in the 1970s, Magnetic Resonance Imaging (MRI) was built on a completely different physical foundation: nuclear magnetic resonance. Instead of X-ray absorption or acoustic reflection, MRI exploited the precession of hydrogen nuclei in a magnetic field and their relaxation after radiofrequency excitation. The image was formed by encoding spatial information through magnetic field gradients. This framework did not compete directly with CT in the sense of replacing it; rather, it offered a radically different contrast mechanism—sensitive to soft tissue, water content, and blood flow—without ionizing radiation. MRI coexisted with CT, each framework dominating different diagnostic domains. The physics of MRI was far more complex than that of CT, involving relaxation times (T1, T2), diffusion, and perfusion, which meant that the subfield had to develop new models of what constituted an image and how to interpret its contrast.
By the 1990s, a methodological rivalry emerged within the reconstruction community. Iterative Reconstruction (IR) challenged the dominance of Filtered Back Projection. Instead of a single analytical formula, IR modeled the image formation process as a system of equations and solved it iteratively, refining the estimate of the image step by step. Its key innovation was that it could incorporate a forward model of the physics—including noise statistics, photon statistics, detector response, and system geometry—directly into the reconstruction. This made IR far more robust to noise and incomplete data than FBP. The cost was computation: early IR was too slow for clinical use. Over the following decades, faster computers and better algorithms (ordered subsets, penalized likelihood) brought IR into clinical practice, first in PET and SPECT, then in CT. The rivalry between FBP and IR was not a simple replacement; IR absorbed many of FBP's insights (the Radon transform, the central slice theorem) while rejecting its assumption that noise could be handled separately. Today, IR has largely replaced FBP in new systems, though FBP remains in use for quality assurance and as a reference standard.
The early 2000s saw the rise of Multimodal Fusion Imaging, a framework that did not introduce new physics but instead coordinated existing physical models. The landmark example was PET/CT: a single scanner that acquired both a PET emission image (functional) and a CT transmission image (anatomical) in the same session. The CT image was used for attenuation correction of the PET data and for anatomical localization. This framework transformed the subfield by treating images from different physical bases as complementary rather than competing. Multimodal Fusion Imaging required new methods for registration, calibration, and joint reconstruction. It did not replace its parent frameworks (Nuclear Medicine Imaging, Computed Tomography, MRI) but created a new layer of infrastructure that combined them. SPECT/CT, PET/MRI, and hybrid ultrasound systems followed. The framework's distinctive claim was that the whole—a fused image—was diagnostically greater than the sum of its parts.
The most recent framework, Deep Learning Reconstruction (DLR) , represents a fundamental shift in how imaging physics thinks about image formation. Instead of modeling the physics explicitly (as IR does) or using an analytical formula (as FBP does), DLR learns the mapping from raw data or low-quality images to high-quality images directly from training examples. The physical model is replaced by a learned model. This framework emerged from advances in deep convolutional neural networks and the availability of large training datasets.
DLR differs from Iterative Reconstruction in a crucial philosophical way. IR says: 'I know the physics, so I can simulate the forward process and invert it.' DLR says: 'I do not need to know the physics explicitly; I can learn the inverse mapping from examples.' In practice, DLR often outperforms IR in speed and image quality, especially at low radiation doses or fast scan times. However, it introduces new problems: generalization to unseen anatomy, sensitivity to training data distribution, and lack of interpretability. The relationship between DLR and IR is one of living disagreement. Some researchers argue that DLR will eventually absorb IR by learning implicit physics models; others insist that physics-based models remain essential for robustness and trustworthiness. Currently, DLR is rapidly being adopted in commercial CT and MRI systems, often combined with IR in hybrid approaches.
Today, imaging physics is a pluralistic field. No single framework has won. Projection Radiography is still used for chest X-rays and mammography. Nuclear Medicine Imaging and Ultrasound Imaging remain essential, each with its own physics and clinical niches. CT and MRI coexist as the dominant tomographic modalities, each with its own reconstruction ecosystem. Within reconstruction, Iterative Reconstruction is the clinical standard for most CT and PET systems, while Deep Learning Reconstruction is the fastest-growing competitor. Multimodal Fusion Imaging has become routine in oncology and neurology.
The leading frameworks today agree on one thing: the goal is to extract the maximum diagnostic information from the minimum radiation dose or scan time. They disagree on how to achieve it. The physics-based camp (IR, model-based reconstruction) argues that explicit modeling of the imaging chain is the only reliable path. The data-driven camp (DLR) argues that learned models can outperform hand-crafted physics models, especially for complex nonlinear effects. A third group seeks synthesis: using deep learning to accelerate or improve iterative reconstruction, combining the strengths of both. This tension—between explicit physical modeling and data-driven learning—is the central intellectual debate in imaging physics today, and it will likely define the next generation of frameworks.