Working Through Medical Imaging Signals And Systems
I spent three semesters teaching this course before I stopped caring enough to keep office hours. The textbook by Parker and Hsieh is decent but the problem sets are where most people fall apart. Students expect the math to be straightforward because it looks like Fourier territory, but then they hit reconstruction filtered backprojection and realize convolution in the frequency domain is not the same thing as convolution in the spatial domain when aliasing is involved. The core signal processing pipeline in medical imaging runs through sampling, reconstruction, and noise analysis. You start with raw projections or k-space data, apply reconstruction algorithms, and end up with pixel values that have no direct physical meaning until you calibrate them against phantom data. Most homework problems skip the calibration step because it is tedious, and that is why your final image always looks slightly wrong. One thing nobody explains well in the textbook is that ramp filtering in CT reconstruction assumes infinitely narrow detector elements. Real detectors have finite apertures, and the system transfer function rolls off at high spatial frequencies. I had a student submit a reconstruction where the edges were razor sharp but the uniform region had a strange oscillating pattern that looked like Gibbs ringing. The actual cause was that he had applied the ramp filter without accounting for the detector window function. The fix was to multiply the ramp by a Hamming window in the frequency domain before inverse transforming. This reduced the edge artifacts at the cost of about 10 percent spatial resolution, which is the tradeoff every CT system makes implicitly.
Why Medical Imaging Signals And Systems Solutions Matter in Practice
The concepts in this course show up directly in modalities you will encounter in any clinical engineering role. CT uses projection geometry and the Radon transform, which maps to filtered backprojection or iterative reconstruction depending on the hardware. MRI works entirely in k-space, and the sampling pattern determines image quality more than the receiver coil sensitivity. Ultrasound beamforming is essentially a delay-and-sum process that can be analyzed with the same signal theory as array processing in radar. Aliasing is the first trap. Nyquist applies to each dimension independently, which means you can undersample the phase encoding direction in MRI and fold the image without affecting the readout dimension. The aliasing artifacts look different in each case. In CT they appear as streaks that rotate with the projection angle. In MRI they create spatial overlap that is often mistaken for pathology. I once spent two days troubleshooting what a radiologist reported as an abnormality, only to find the phase encoding sampling rate was insufficient for the field of view selected. The solution was adjusting the matrix size or applying parallel imaging with SENSE or GRAPPA instead of just skipping the step entirely. Noise characterization is another area where the textbook oversimplifies. Gaussian noise models work fine for CT when the photon count is high, but low dose studies introduce Poisson statistics that dominate the noise texture. MRI noise follows a Rician distribution in magnitude images, which means the background is never truly zero even when there is no signal. If you compute noise power spectrum from a region of interest near the image edge, you will overestimate the noise level because the Rician bias inflates the apparent standard deviation. The workaround is to use complex difference methods or acquire a noise reference scan with the RF transmitter turned off.
Signal to noise ratio calculations in this course typically assume additive white Gaussian noise, which is convenient but rarely accurate. In reality, spatially correlated noise from reconstruction filters and temporal noise from patient motion create different degradation patterns. A practical improvement is to estimate the noise power spectrum from a uniform phantom and design your filter based on the measured NPS rather than the theoretical one. This takes about 20 minutes per scanner type but saves hours of trial and error during protocol optimization.
Get the Full Details
Core Methods You Need to Actually Use
Filtered backprojection remains the standard reconstruction method for fan beam and cone beam CT because it is fast enough for clinical real-time use. The Mathews kernel or Shepp-Logan filter handles the ramp filtering, and you choose between them based on whether you prioritize resolution or noise suppression. The Ram-Lak filter gives the sharpest image but amplifies high frequency noise to unusable levels in low dose scans. Shepp-Logan introduces a sinc apodization that reduces noise without significant resolution loss, which is why most commercial systems default to it. K-space sampling in MRI deserves more attention than it gets in typical coursework. Cartesian sampling is straightforward but slow. Radial and spiral sampling fill k-space differently and produce distinct artifact patterns. Radial sampling creates streak artifacts instead of aliasing folds, which are often easier to recognize and tolerate. Spiral sampling is efficient but requires gridding reconstruction because the data does not fall on a rectilinear grid. The gridding process introduces a density compensation step that if skipped creates a bright center and dark periphery in the reconstructed image. Iterative reconstruction is gaining ground in CT and PET, replacing analytic methods in many new systems. The basic approach involves initializing an image estimate, forward projecting it through the system model, comparing the result with measured data, and updating the image using a regularization term. The regularization parameter controls the tradeoff between data fidelity and noise suppression. Setting this parameter too low produces noisy images that look acceptable to a radiologist only because they are familiar with noise patterns. Setting it too high produces smooth images that obscure low contrast lesions. I calibrate this parameter once per scanner using a standard phantom at representative dose levels and reuse those values rather than retuning it for every new protocol.
Image quality metrics matter more than grades on these problem sets. The modulation transfer function describes spatial resolution degradation across frequencies. The noise power spectrum describes how noise distributes across spatial frequencies. Detectability index combines both into a single measure of task performance. Most homework assignments ask for MTF or NPS separately because that is easier to grade. In practice you need both to evaluate a reconstruction method properly. If someone claims their algorithm improves resolution, ask for the MTF and the NPS together. A resolution gain that comes with a noise penalty may not improve detectability at all.
Common Pitfalls and What to Do Instead
The biggest mistake students make is confusing spatial domain operations with frequency domain operations. Convolution in the spatial domain equals multiplication in the frequency domain only under specific conditions that are usually violated in imaging. Circular convolution introduces wrap-around artifacts that corrupt the image edges. Padding the input with zeros eliminates this problem but increases computation time. For 512 by 512 images, adding two rows and columns of zeros typically increases memory usage by less than one percent while preventing edge corruption entirely. Another common error is applying filtering without considering the sampling density. If you design a filter for one sampling interval and apply it to data sampled at a different rate, the frequency response shifts. This happens frequently when researchers compare reconstruction methods across different manufacturers using different native resolutions. Always resample to a common grid before comparing filter responses or measuring image quality metrics. Window functions deserve careful selection. The choice between Hamming, Hanning, and Blackman affects both resolution and noise suppression differently. Hamming provides good noise suppression with moderate resolution loss. Hanning is similar but with slightly better resolution. Blackman provides the strongest noise suppression but the largest resolution penalty. In cardiac CT where motion is present, you might prefer Hanning to preserve as much resolution as possible. In abdominal CT with low contrast lesions, Blackman might be worth the resolution cost because noise reduction improves detectability more than resolution gain.
The discrete nature of digital imaging introduces quantization effects that are often ignored. Pixel values from CT are stored as 12 or 16 bit integers mapped to Hounsfield units. The quantization step size determines the smallest detectable intensity change. For a 12 bit system with a 4000 HU range, each step is about 1 HU. This is sufficient for most clinical tasks but becomes limiting when performing quantitative analysis like bone mineral densitometry where sub-HU changes matter. The solution is to retain the full bit depth during processing and quantize only at the final display stage.
What This Approach Cannot Do Well
Signals and systems theory breaks down when dealing with nonlinear phenomena. Metal artifacts in CT arise from beam hardening and scattering, which are nonlinear effects that cannot be modeled with linear filter theory. Iterative metal artifact reduction algorithms exist but they are empirical rather than derived from first principles. Similarly, partial volume effects in MRI and CT create intensity averaging that violates the linear assumption underlying most reconstruction methods. These limitations mean that no amount of frequency domain analysis will fully correct these artifacts. You need physical models or empirical corrections instead. The theory also assumes stationary systems, which is rarely true in practice. Scanner gain drifts over time. Patient position changes between scans. Detector elements age at different rates. A filter designed for ideal conditions degrades when applied to real data. Regular recalibration is necessary, and the frequency at which it must be done depends on the scanner type and usage intensity. PET scanners require daily normalization due to detector gain variations. CT scanners benefit from weekly calibration but may tolerate longer intervals if operated in a stable environment. Computational requirements scale poorly with image size. A full 3D filtered backprojection reconstruction for a modern CT scanner with 1000 projections and a 1024 by 1024 by 512 volume requires roughly 10 to 15 seconds on a standard workstation. Iterative reconstruction for the same data can take 30 minutes to several hours depending on the number of iterations and regularization complexity. This is why most clinical systems use hybrid approaches that combine analytic initialization with a small number of iterative refinement steps.
If you are looking for structured worked examples and detailed derivations, most universities distribute solution manuals through their course websites or textbook publishers. The solutions for Parker and Hsieh cover the standard problem sets including convolution integrals, Fourier transforms of projection data, and reconstruction filter design. Working through these examples manually before checking the answers is the most reliable way to build intuition, though the process typically takes 2 to 3 hours per problem set for someone with a solid mathematics background. The practical takeaway is that signals and systems theory gives you the foundation but clinical imaging requires adapting that theory to nonideal conditions. Learn the derivations, understand the assumptions, and then learn which assumptions break in practice. That gap between theory and practice is where most of the useful knowledge actually lives.
