Essential Physics Of Medical Imaging

Why Your CT Reconstructions Look Wrong

The physics behind medical imaging isn't particularly complicated, but the engineering layer on top of it introduces enough noise and artifacts that most people working in the field don't fully understand what they're looking at. I've spent years debugging reconstruction pipelines and calibrating imaging systems, and the number one mistake I see is people treating the raw data like it tells the whole story. It doesn't. Let's start with how a photon actually behaves when it hits matter. That's where everything begins. A photon traveling through tissue can pass straight through with no interaction, get completely absorbed via the photoelectric effect, or scatter at an angle through Compton scattering. The probability of each event depends on the photon energy and the atomic number of the material. In CT, we're typically working in the 30 to 140 keV range, and in that window both Compton and photoelectric effects are fighting for dominance. Bone absorbs differently than soft tissue. Iodine contrast changes the absorption curve in a way that matters for image quality. Understanding this is the foundation of the Essential Physics Of Medical Imaging because everything else builds on it.

Essential Physics Of Medical Imaging

The Beer-Lambert law is your starting point. It describes exponential attenuation: the intensity coming out of a material is the initial intensity multiplied by e to the negative power of the linear attenuation coefficient times the path length. In practice, that coefficient isn't a constant. It varies with energy, which is why dual-energy CT exists. You're essentially taking two measurements at two different energy spectra and solving for material composition rather than just density. The linear attenuation coefficient for water at 80 keV is roughly 0.18 cm^-1. For cortical bone at the same energy it's closer to 0.5 cm^-1. Soft tissue sits somewhere in between. These numbers shift dramatically when you change the kVp setting. Dropping from 140 kVp to 80 kVp nearly doubles the photoelectric contribution for iodine, which is why low-kVp protocols are useful for angiography but absolutely wreck image quality in larger patients. The scatter increases, the photon starvation becomes real, and your noise goes through the roof. I remember running into this specific problem a few years back. We were doing contrast-enhanced CT on pediatric patients at 80 kVp because the protocol book said it would maximize iodine contrast. What the book didn't say is that for kids under 15 kilograms, the photon flux at that energy was so low after passing through even thin bodies that the reconstructed images were basically unusable noise. We ended up switching to 100 kVp with a stronger iodine dose and an iterative reconstruction algorithm. The contrast-to-noise ratio improved by about 40 percent compared to the original protocol, and the diagnostic quality was actually usable. The takeaway is that protocol optimization is never just about picking the highest contrast setting. You have to balance photon starvation, scatter, and patient size simultaneously.

Detectors and Signal Chain

The detector is where the X-rays actually become something you can process. Modern CT scanners use solid-state scintillator detectors, typically gadolinium oxysulfide or cadmium tungstate, coupled to photodiodes. The scintillator converts X-ray photons into visible light, and the photodiode converts that light into an electrical charge. The whole chain has to be fast, efficient, and stable. Any drift in the detector response and your Hounsfield units will shift, which means your liver and your spleen won't look where they're supposed to look on the calibrated scale. Detector efficiency is usually in the 70 to 90 percent range for diagnostic energies, but that efficiency isn't uniform across the entire detector array. Individual detector elements degrade at different rates over time. I've calibrated scanners where a single bad detector row was causing streak artifacts that looked like beam hardening but weren't. The fix was replacing the detector module. Without understanding the physics of how detector response maps to image artifacts, you could spend days chasing reconstruction parameters before ever checking the hardware. CT numbers exist because we normalize everything to water. Water is defined as zero Hounsfield units and air is negative one thousand. Dense materials like cortical bone or metal implants sit well above zero. The relationship is linear only within the calibrated range, and outside that range things get messy. Metal implants produce phantoms and streaks that have nothing to do with the actual anatomy and everything to do with photon starvation and scatter. That's the physics fighting the engineering.

MRI Physics: Larmor Precession and Relaxation

MRI operates on an entirely different physical principle. There are no X-rays involved. Instead, you're manipulating the magnetic moments of hydrogen nuclei, mostly in water and fat molecules, using strong static magnetic fields and radiofrequency pulses. The fundamental frequency is the Larmor frequency, which is proportional to the magnetic field strength. At 1.5 Tesla, that's about 64 megahertz for hydrogen. At 3 Tesla, it doubles to roughly 128 megahertz. Higher field strength gives you better signal-to-noise ratio, but it also amplifies certain artifacts and makes dielectric shading more noticeable in the body coil. The relaxation times T1 and T2 are what make MRI useful. T1 is the longitudinal recovery time, how quickly the net magnetization vector realigns with the main magnetic field after being tipped. T2 is the transverse decay time, how quickly the spins dephase due to interactions with neighboring spins. T2* includes additional dephasing from magnetic field inhomogeneities. These times vary significantly between tissue types. Gray matter has a longer T2 than white matter. Fat has a short T1. Water has a very long T1 and T2. Pulse sequence design is essentially the art of weighting the image contrast to emphasize the difference between the tissues you care about. A gradient echo sequence with a short TR and short TE gives T1 weighting. A spin echo with long TR and long TE gives T2 weighting. Proton density weighting uses long TR and short TE to minimize both contrast mechanisms. This is standard textbook stuff, but the practical implications are where things get interesting. When I was working on a project to optimize MR lymphography, I found that the standard T2-weighted sequences were missing small lymph nodes because the contrast between the node and surrounding fat wasn't sufficient. Switching to a Dixon-based fat-suppressed sequence increased the conspicuity of sub-centimeter nodes by a measurable margin. The physics hadn't changed, but the sequence design was exploiting the relaxation differences more effectively.

K-Space and Image Formation

K-space is the Fourier domain representation of the MR image. Every point in k-space contains information about the entire image, which is the opposite of how CT projection data works. Filling k-space centrally gives you the broad contrast and low-frequency anatomy. The periphery contains the fine detail and edges. Partial k-space acquisition shortcuts the process by acquiring slightly less than half the data and relying on conjugate symmetry to reconstruct the rest. You save time but introduce blurring if the symmetry assumption is violated by motion or flow. Parallel imaging uses multiple receiver coils with different sensitivity profiles to undersample k-space. SENSE and GRAPPA are the two main approaches. The acceleration factor determines how much you skip, and higher acceleration factors increase the g-factor noise penalty. At an acceleration factor of 3, the SNR drop can be substantial in the center of the coil array, which is why you often see higher acceleration factors used in extremity imaging where coil geometry is favorable and lower factors in abdominal imaging where coverage is more challenging. Ultrasound physics is another universe entirely. You're sending mechanical pressure waves into tissue and listening for echoes. The impedance mismatch between tissues determines reflection strength. Bone reflects almost everything, which is why ultrasound can't image through the skull in adults. Air reflects almost everything too, which is why lung imaging with ultrasound is limited to pleural abnormalities. Soft tissue interfaces create the gray-scale image you're familiar with.

The speed of sound in soft tissue is assumed to be 1540 meters per second in most scanners. That assumption breaks down in fat, where the speed is closer to 1450 m/s, and in muscle, where it's closer to 1580 m/s. The mismatch causes refraction artifacts and focusing errors, especially at higher frequencies. Frequency selection is always a tradeoff. Higher frequency gives better resolution but less penetration. A 12 MHz linear probe might resolve structures at 3 centimeters depth, but a 5 MHz curvilinear probe will reach 15 centimeters with acceptable signal. You pick the probe based on what you're trying to see, not what sounds best on paper.

Image Reconstruction Algorithms

Filtered back projection was the standard for decades. It's fast, deterministic, and straightforward. You take the raw projections, apply a ramp filter to correct for the blurring inherent in simple back projection, and sum everything back into image space. The problem is that FBP assumes perfect data. Real data has noise, artifacts, and incomplete sampling. When you reduce the radiation dose in CT, FBP images become grainy quickly because the algorithm doesn't have a mechanism to distinguish signal from noise. Iterative reconstruction methods changed that landscape. They model the imaging physics forward, compare the simulated projection to the actual measured data, and update the image estimate repeatedly until convergence. This takes more compute time but produces cleaner images at lower doses. Most modern scanners offer some form of iterative reconstruction as standard. The tradeoff is that iterative methods can introduce a slightly smoother, sometimes plasticky appearance to the image texture. Radiologists need to adapt to that, and some initially resist it because the texture doesn't match what they trained their eyes on. Deep learning-based reconstruction is now entering clinical practice. Networks trained on paired low-dose and standard-dose data can reconstruct images that look like standard-dose scans even when the input data is noisy or sparse. The performance is impressive, but the generalization is the concern. A model trained on data from one scanner manufacturer at one institution may not translate well to a different scanner or a different patient population. I saw a case where a vendor's AI reconstruction tool produced artifacts at bone-soft tissue interfaces that weren't present in the raw data. The network was hallucinating structure to fill in missing information, and while the images looked better visually, they contained information that wasn't actually there. That's a real risk with black-box reconstruction methods.

Nuclear Medicine and PET

Positron emission tomography relies on a completely different physics pathway. You inject a radiotracer labeled with a positron-emitting isotope, most commonly fluorine-18. The positron travels a short distance in tissue, annihilates with an electron, and produces two 511 keV gamma photons traveling in opposite directions. The PET scanner detects these coincident photon pairs and localizes the annihilation event along the line of response. The spatial resolution of PET is limited by several factors. The positron range before annihilation blurs the localization. Non-collinearity of the two photons, a deviation of about 0.5 degrees from perfect 180-degree separation, adds uncertainty. The detector crystal size sets a fundamental limit on how precisely you can localize the event. Modern PET scanners with smaller crystals and time-of-flight capability achieve resolutions around 4 to 5 millimeters, which is decent but nowhere near the sub-millimeter resolution of CT or MRI. Quantification in PET uses standardized uptake values, which normalize the measured activity concentration by the injected dose and patient body weight. SUV is useful for tracking treatment response over time in oncology, but it's sensitive to timing, blood glucose levels, and body composition. A patient with high blood glucose will have altered tracer distribution, and an obese patient will have a different normalization factor than a lean patient. The physics is straightforward, but the clinical application requires attention to pre-scan preparation and standardization protocols. I once reviewed a PET scan from a facility that wasn't standardizing the fasting time before injection, and the variability in SUV values was so large that therapeutic response assessment was essentially meaningless.

Practical Considerations That textbooks Ignore

The physics is clean. The practice is messy. Beam hardening in CT isn't just a correction you apply once during calibration. It changes with patient size, with the presence of contrast material, and with the spectral shape of the X-ray tube. Automatic exposure control modulates the tube current based on patient attenuation in real time, which means the spectrum and intensity vary across the projection angles. This variation is what creates the cone-beam artifacts in wide-detector CT scanners when the anatomy extends beyond the standard reconstruction field. Aliasing in MRI is another issue that doesn't get enough attention in introductory courses. If the field of view is smaller than the object being imaged, signals from outside the FOV fold back into the image. Anti-aliasing techniques like oversampling in the phase-encode direction or using selective excitation pulses can mitigate this, but they add complexity and sometimes reduce efficiency. In practice, I've seen aliased MRIs misdiagnosed as anatomical anomalies because the person interpreting the scan didn't recognize the folding pattern. Ionizing radiation dose is the most discussed safety concern in medical imaging, and it deserves the attention. The effective dose from a typical chest CT is around 7 millisieverts, which is roughly equivalent to two and a half years of natural background radiation. The risk is stochastic, meaning there's no threshold below which the risk is zero, but the absolute risk from a single scan is small. The benefit-risk calculation depends on the clinical question. A CT for acute trauma in an emergency department is a different equation than a CT for routine surveillance in an asymptomatic patient.

MRI has no ionizing radiation, but it introduces its own set of physical constraints. The acoustic noise from gradient switching can exceed 120 decibels, which is above the threshold for potential hearing damage without protection. Patients need earplugs or headphones. The strong magnetic field means any ferromagnetic object in the room becomes a projectile. Implant screening is non-negotiable. Someone with an older aneurysm clip or a cochlear implant cannot go into a 3 Tesla scanner. The physics of magnet interaction is unforgiving.

What Actually Matters in Practice

Understanding the physics helps you make better decisions about protocol selection, artifact identification, and quality control. It doesn't replace the need for, but it gives you a framework for understanding why an image looks the way it does. When you see a streak artifact, you can ask whether it's beam hardening, motion, metal, or detector failure. When an MRI sequence produces unexpected contrast, you can trace it back to the TR, TE, and flip angle choices and predict how changing them will affect the outcome. The field is moving toward hybrid imaging and AI-assisted reconstruction. PET/MRI combines two modalities into a single system, eliminating the registration errors that come from separate scans. AI methods are being applied to dose reduction, artifact suppression, and automated segmentation. These tools are powerful, but they operate on the same physical principles. The physics doesn't change because the software gets smarter. Understanding those principles remains the most reliable way to troubleshoot problems, validate results, and communicate with physicists and engineers who build and maintain these systems.