Filtered Back Projection (FBP)

FBP is the direct practical implementation of the Fourier Slice Theorem (also known as the Central Slice Theorem). This theorem states that the 1D Fourier Transform of a parallel projection at a specific angle is exactly equal to a radial line through the 2D Fourier Transform of the original object at that same angle. FBP reconstructs the image by 'filling in' the 2D frequency domain with these radial samples and then performing an inverse 2D Fourier Transform. The filtering step corrects for the non-uniform sampling density in the frequency domain, which is denser at the center (low frequencies) than at the periphery (high frequencies).

The 'filtering' in FBP is a mathematical convolution applied to each projection before back-projection. The core of this filter is a ramp function (or Ram-Lak filter), which is a high-pass filter in the frequency domain. Its purpose is to counteract the inherent 1/r blurring that occurs during simple back-projection. Without it, the image would be an unusable, heavily blurred version of the original object. In practice, the ramp filter is often combined with a low-pass apodization window (e.g., Hamming, Hann, Shepp-Logan) to suppress the amplification of high-frequency noise, trading off spatial resolution for a smoother image.

FBP's primary historical advantage is its computational efficiency. It is a linear, non-iterative algorithm that can reconstruct a slice in a fraction of a second, making it the only viable option for clinical CT from the 1970s until the 2010s. Its linearity also means that image quality properties, such as noise texture and resolution, are highly predictable and well-understood. However, this linearity is also a key limitation: FBP cannot model non-linear physical effects like beam hardening, scatter, or photon starvation, which are better handled by modern iterative reconstruction (IR) and deep learning reconstruction (DLR) methods.

A fundamental weakness of FBP is its non-local noise propagation. A single noisy or corrupted detector measurement in the sinogram is smeared along the entire back-projection path, creating characteristic streak artifacts across the image. This is particularly problematic in low-dose CT scans, where photon starvation leads to severe noise amplification by the ramp filter. FBP lacks any mechanism to regularize or locally adapt to noise, resulting in a direct trade-off between radiation dose and image noise. This limitation is the primary driver behind the industry's shift to iterative and deep learning-based reconstruction.

The classic FBP derivation assumes a parallel-beam geometry, where each projection consists of a set of parallel X-ray paths. Modern clinical scanners use a fan-beam geometry with a divergent X-ray source. To apply FBP, fan-beam data must be rebinned or re-sorted into equivalent parallel projections. Alternatively, a modified fan-beam FBP algorithm can be used, which incorporates a cosine weighting factor and a distance-dependent back-projection term. Cone-beam geometries, common in modern multi-detector CT, require even more sophisticated approximations like the Feldkamp-Davis-Kress (FDK) algorithm, which is an approximate 3D extension of FBP.

The final step of FBP is back-projection, which is the mathematical inverse of the forward projection (Radon transform). For each filtered projection at a given angle, the intensity values are 'smeared' back across the entire image grid along the original ray paths. The final reconstructed image is the superposition of these smeared projections from all acquisition angles. This process is geometrically intuitive: a single point in the object contributes to multiple projections, and back-projection correctly redistributes that contribution. The filtering step ensures that this redistribution is mathematically exact, not just an approximation.

The mathematical foundation of FBP. The Radon transform maps a 2D function (the image) into a set of line integrals (the projection data). Each point in the sinogram represents the integral of attenuation along a specific ray path through the object. FBP is essentially an analytic inversion of this transform.

• The Central Slice Theorem links the Radon transform to the Fourier domain, proving that a 1D Fourier transform of a projection equals a line through the 2D Fourier transform of the object.
• FBP exploits this theorem by filtering projections in the frequency domain before back-projection.

The high-pass frequency filter applied to projection data before back-projection. Without it, simple back-projection produces a severely blurred image with a characteristic 1/r point spread function.

• The Ram-Lak filter has a frequency response linearly proportional to |ω|, exactly compensating for the oversampling of low frequencies during back-projection.
• In practice, the ramp is often apodized (windowed) with functions like Shepp-Logan, Hamming, or Hann to trade off spatial resolution for noise suppression.
• The choice of kernel directly controls the noise texture and edge sharpness in the final image.

The modern successor to FBP in diagnostic CT. Unlike FBP's single-pass analytic approach, IR algorithms model the acquisition physics—including focal spot geometry, detector response, and photon statistics—and iteratively refine the image estimate.

• Model-Based IR (MBIR) incorporates system optics and noise statistics, achieving dramatic noise reduction at the cost of long reconstruction times.
• Hybrid IR blends FBP with statistical denoising in image or projection space, offering a practical balance of speed and image quality.
• IR excels in low-dose protocols where FBP images become unacceptably noisy, making it the standard for pediatric and screening CT.

The latest evolution in CT reconstruction, using convolutional neural networks trained to map low-quality FBP or IR images to high-quality, low-noise outputs. DLR overcomes the unnatural noise texture and plastic-like appearance sometimes associated with aggressive IR settings.

• Networks are trained on paired low-dose and routine-dose images, or using self-supervised methods on routine-dose data.
• DLR can resolve fine anatomical structures—such as trabecular bone and coronary artery plaques—that are obscured by noise in conventional reconstructions.
• Commercial implementations include Canon's AiCE, GE's TrueFidelity, and Siemens' ReconCT, each using proprietary deep learning architectures.

A fundamental limitation of approximate FBP in cone-beam geometry, where the X-ray source diverges in both the axial and longitudinal directions. The standard Feldkamp-Davis-Kress (FDK) algorithm is an approximate FBP extension that becomes increasingly inaccurate with larger cone angles.

• Cone-beam artifacts manifest as shading, streaking, and geometric distortion, particularly severe in peripheral regions of wide-detector CT scans.
• Modern 256-slice and 320-slice scanners with wide z-axis coverage require exact or iterative reconstruction to mitigate these errors.
• Exact analytic solutions, such as the Katsevich algorithm, exist for helical trajectories but are computationally complex.

A critical post-processing challenge where metallic implants—such as hip prostheses, dental fillings, and surgical clips—cause severe photon starvation and beam hardening in projection data. FBP faithfully reconstructs these corrupted projections as bright and dark streaks.

• Projection-interpolation MAR replaces corrupted sinogram data with interpolated values from neighboring projections, but can introduce new artifacts.
• Normalized MAR (NMAR) improves on this by performing interpolation in a normalized domain after segmenting the metal in a prior image.
• Deep learning MAR uses neural networks to directly inpaint corrupted sinogram regions or correct the reconstructed image, often trained on paired artifact-free and artifact-simulated data.