A CT scanner does not capture the final image directly
The scanner acquires projection measurements from many angles around the subject. Those measurements live in the projection domain, commonly represented as a sinogram, not as the anatomical slice a radiologist eventually reads.
This point is basic, but it matters for almost every reconstruction method on the site. Some models operate after reconstruction in image space. Others work in the measurement domain before the final image exists. The lab’s self-supervised work in particular depends on taking that distinction seriously.
Reconstruction is an inverse problem
Once the measurements are collected, the task is to recover the spatial structure that produced them. That is the inverse problem at the center of CT. Given line-integral observations from many angles, estimate the attenuation map inside the body.
This is why CT reconstruction is more than image enhancement. The algorithm is not merely cleaning an image. It is inferring one from incomplete and noisy evidence.
Why filtered back-projection remains the baseline
Filtered back-projection is still everywhere because it is fast and direct. It filters the measurements and back-projects them into image space in one analytical pipeline. When the acquisition quality is high, the method works remarkably well.
Its weakness is equally well known: it has no explicit notion of uncertainty or prior structure. If the measurements are sparse or noisy, the method still produces an answer quickly, but that answer may contain streaks, amplified noise, or missing low-contrast detail.
What iterative reconstruction changes
Iterative methods replace one-shot inversion with repeated refinement. Start with an estimate, simulate what measurements that estimate would generate, compare the result with the real data, and update the image. Repeat until the estimate becomes more consistent with the acquisition and the chosen prior.
ART, SIRT, and SART differ in how aggressively and how often they update the image, but the underlying idea is the same: quality improves when the reconstruction checks itself against the measurement model instead of inverting once and stopping.
Why regularization matters so much
Without regularization, iterative reconstruction can still overfit noise. Regularization adds a belief about what a plausible image should look like. Total variation is the canonical example because it favors piecewise smooth regions while preserving edges.
Many modern reconstruction papers can be read as new answers to a very old question: what prior should the algorithm trust? Handcrafted priors, learned priors, and hybrid priors are all different answers with different tradeoffs in speed, interpretability, and data requirements.
Why this stack matters for the lab’s papers
This is the context behind the papers on the site. Some improve the image after reconstruction. Some operate in the projection domain. Some redesign the loss, and some redesign the prior. The common thread is that CT reconstruction starts from acquisition and inverse problems before it becomes a machine learning question.