CT Reconstruction: Radon Transform, Fourier Slice Theorem, & Convolution Backprojection

Added:

Iterative Reconstruction and Deep Learning-Based Methods

While analytical methods like Filtered Backprojection (FBP)—founded on the Radon transform and Fourier Slice Theorem—are computationally efficient, they perform poorly under low-dose or sparse-view conditions, leading to severe noise and streak artifacts. The primary counter-paradigm is Iterative Reconstruction (IR) and, more recently, Deep Learning-based reconstruction. Unlike FBP's direct mathematical inversion, IR formulates reconstruction as an optimization problem, incorporating physical models of the X-ray system, noise statistics, and prior image constraints (such as total variation). This allows for high-quality image generation from significantly noisier or incomplete projection data, drastically reducing the patient's radiation dose. Furthermore, deep learning models can learn complex, non-linear mappings directly from projections to image space or refine analytical outputs, challenging the necessity of strict adherence to classical analytical transform theories in modern CT imaging.

Fundamentals of X-ray physics, including attenuation, the Beer-Lambert law, and how 2D projection data (sinograms) are acquired.
The Fourier Transform, specifically 1D and 2D continuous Fourier transforms and the frequency/spatial domain relationship.
Linear systems theory, including the concept of convolution, filtering, and the Convolution Theorem.
Multivariable calculus and coordinate systems, particularly line integrals and converting between Cartesian and polar coordinates.
Iterative Reconstruction techniques (such as ART, SART, and MLEM) used to improve image quality under low-dose or sparse-view conditions.
Advanced CT scan geometries, such as Fan-Beam and Cone-Beam reconstruction algorithms (e.g., the FDK algorithm).
Analysis and mitigation of CT image artifacts, including beam hardening, metal artifacts, and motion blur.
Deep learning-based CT reconstruction methods, such as learned primal-dual reconstruction and convolutional neural networks for sinogram denoising.
27.4K views372likes19:45@jpacademia4716Original Release: 2020-10-30

CT reconstruction uses mathematical tools to convert 2D projection images from multiple angles into 3D anatomical representations. The Radon transform mathematically describes how projections are formed by integrating the object's attenuation coefficient along lines at specific angles. The Fourier Slice Theorem establishes that the 1D Fourier transform of a projection equals a radial slice through the 2D Fourier transform of the object at the same angle, forming the theoretical basis for reconstruction. However, practical implementation requires convolution backprojection, which applies a high-pass filter (such as a ramp filter) to the projection data before performing backprojection to correct for blurring and reconstruct the original image efficiently.