Filtered Back Projection and Reconstruction Kernels

Analytical Reconstruction: Simple Back Projection & The $1/r$ Star Artifact

Reconstruction is the mathematical procedure that converts projection ray sums ($p$) into an array of linear attenuation coefficients ($\mu$) within the pixel matrix.

Simple (Unfiltered) Back Projection Mechanics

The most intuitive analytical method is simple back projection (summation method). In simple back projection, the computer takes the measured attenuation profile obtained at projection angle $\theta$ and projects it uniformly back along straight lines across the digital matrix. Each pixel along a ray's path is credited with an equal fraction of the total measured attenuation.

This process is repeated sequentially for all angular projections acquired between $0^\circ$ and $360^\circ$, with the intersecting back-projected rays summing their values in each pixel cell:

$$f_{\text{backprojected}}(x,y) = \int_0^\pi p(x\cos\theta + y\sin\theta,\, \theta)\, d\theta$$

The Mathematical Limitation: The $1/r$ Blurring & Star Artifact

While simple back projection reproduces the general location of structures, it suffers from a fundamental mathematical flaw: spurious density accumulation. Because the projection rays are smeared entirely across the matrix, density is assigned not only to the true object but also to every non-attenuating pixel lying along the back-projection path.

When hundreds of angular projections are summed, these overlapping smearing rays produce a characteristic point spread function that decays inversely with radial distance ($r$) from the high-contrast point:

$$PSF(r) \propto \frac{1}{r}$$

This phenomenon manifests visually as:

• A severe diffuse haze or halo surrounding every structure.
• Radiating spokes or streak lines known as the star artifact.
• Substantial degradation of high-contrast edges and total obliteration of low-contrast soft-tissue interfaces.

Because of this $1/r$ blur, simple back projection cannot produce diagnostically acceptable clinical images.

---

Filtered Back Projection (FBP) Mechanics

Filtered Back Projection (FBP) is the analytical algorithm that solved the $1/r$ blurring problem, serving as the foundational imaging standard for computed tomography for four decades.

The Projection-Slice Theorem (Central Slice Theorem)

FBP is rooted mathematically in the Projection-Slice Theorem (also known as the Fourier Slice Theorem). This theorem states that:

The one-dimensional Fourier transform of a parallel projection of an object at angle $\theta$ is mathematically identical to a central radial slice passing through the origin of the two-dimensional Fourier transform of the object at the identical angle $\theta$.

In the frequency domain, the $1/r$ blurring corresponds to an excessive accumulation of low-spatial-frequency data at the origin of the 2D Fourier space (where all 1D radial projection lines intersect and overlap). To correct this low-frequency over-representation, the projection profiles must be filtered before they are back-projected.

The Two Computational Approaches to FBP

Reconstruction computers implement FBP through one of two mathematically equivalent pathways:

1. Frequency Domain Approach (Fast Fourier Transform):
   • Step 1: Compute the 1D Fourier Transform of each measured projection profile $p(r,\theta)$, converting spatial attenuation profiles into spatial frequency spectra $P(\omega,\theta)$.
   • Step 2: Multiply the frequency spectrum $P(\omega,\theta)$ by a mathematical filter function $H(\omega)$ (the ramp filter, $|\omega|$).
   • Step 3: Compute the Inverse 1D Fourier Transform of the filtered spectrum to return to the spatial domain.
   • Step 4: Back-project the filtered projections across the image matrix.
2. Spatial Domain Approach (Convolution):
   • Rather than converting data into the frequency domain and back via Fourier transforms, the computer performs direct mathematical convolution in the spatial domain.
   • Each raw projection profile is convolved with a pre-calculated discrete reconstruction filter (the spatial impulse response of the filter, $h(r)$):

$$p_{\text{filtered}}(r,\theta) = p(r,\theta) * h(r) = \int_{-\infty}^{\infty} p(\tau,\theta)\, h(r - \tau)\, d\tau$$

• The filtered projection profiles are then back-projected across the matrix.

The Ram-Lak (Ramp) Filter

The ideal mathematical filter that perfectly cancels the $1/r$ blur is the ramp filter (Ramachandran-Lakshminarayanan, or Ram-Lak filter). In frequency space, its transfer function is directly proportional to spatial frequency:

$$H_{\text{ramp}}(\omega) = |\omega|$$

The ramp filter sets zero weight to the DC component (zero frequency), provides low weighting to low spatial frequencies, and linearly amplifies high spatial frequencies. This cancels out the excess low-frequency density accumulation, completely suppressing the $1/r$ star artifact and restoring razor-sharp boundary definition.

---

Reconstruction Algorithms & Kernels: Smoothing vs. Sharp Filters

While a pure ramp filter reduces $1/r$ blurring, its linear amplification of high spatial frequencies introduces a severe drawback: it amplifies high-frequency quantum mottle (statistical photon noise) dramatically. Real-world diagnostic CT therefore modifies the basic ramp filter by multiplying it with a window function (e.g., Hann, Hamming, Shepp-Logan, or Butterworth windows) to create specialized reconstruction kernels (also referred to as reconstruction algorithms or filters).

The selection of reconstruction kernel is the single most powerful software parameter governing the trade-off between spatial resolution and contrast resolution (noise).

Standard / Soft Tissue Kernels (Smoothing Filters)

• Frequency Response: These kernels roll off or cut off high spatial frequencies, attenuating frequencies beyond a moderate threshold.
• Impact on Image Quality: Suppresses quantum mottle and image graininess. Lowering image noise markedly improves contrast resolution (the ability to distinguish between adjacent tissues that have very subtle differences in attenuation, such as gray matter vs. white matter, or a hypodense liver metastasis vs. normal liver parenchyma).
• Trade-off: Blurs high-contrast edges and fine structural margins, reducing high-contrast spatial resolution.
• Primary Indications: Brain CT (evaluating acute ischemic stroke and edema), routine chest (mediastinal window), routine abdomen and pelvis (liver, spleen, pancreas, kidneys, bowel wall).

Bone / Sharp / High-Frequency Kernels (Edge-Enhancing Filters)

• Frequency Response: These kernels maintain high filter gains out to the Nyquist frequency limit, accentuating high spatial frequencies.
• Impact on Image Quality: Greatly sharpens structural boundaries, interfaces, and fine lines, maximizing spatial resolution (the ability to resolve small, closely spaced objects of high contrast).
• Trade-off: Markedly amplifies quantum mottle. Soft-tissue parenchymal evaluation is severely compromised by excessive image graininess.
• Primary Indications: High-resolution CT (HRCT) of the chest (pulmonary interstitium, interlobular septa, bronchiectasis), petrous temporal bones (middle and inner ear ossicles, cochlea, semicircular canals), facial bones, and musculoskeletal trauma (cortical bone microfractures).

---

Step-by-Step Worked Calculations & Practice Scenarios

Worked Example 1: Calculating Pixel Dimension and Voxel Volume

Clinical Scenario: A technologist reconstructs a high-resolution temporal bone scan using a $1024 \times 1024$ matrix, a Display Field of View (DFOV) of $120\text{ mm}$ ($12\text{ cm}$), and a nominal reconstructed slice thickness of $0.5\text{ mm}$.

The pixel dimension is 120/1024 = 0.1171875 mm. Pixel area is about 0.01373 mm², and nominal voxel volume is 0.01373 × 0.5 = 0.00687 mm³. These dimensions describe sampling. The unequal in-plane and longitudinal dimensions are not cubic, and choosing a still smaller pixel does not establish a corresponding physical resolution.

Worked Example 2: Impact of DFOV on Spatial Resolution and Pixel Area

Clinical Scenario: A routine abdominal scan is originally reconstructed using a $400\text{ mm}$ DFOV on a $512 \times 512$ matrix. The radiologist requests a targeted reconstruction of the right adrenal gland using a $150\text{ mm}$ DFOV on the same $512 \times 512$ matrix.

• Step 1: Calculate Original Pixel Size and Area:

$$d_{\text{orig}} = \frac{400\text{ mm}}{512} \approx 0.781\text{ mm} \implies \text{Area}_{\text{orig}} = (0.781)^2 \approx 0.610\text{ mm}^2$$

• Step 2: Calculate Targeted Pixel Size and Area:

$$d_{\text{target}} = \frac{150\text{ mm}}{512} \approx 0.293\text{ mm} \implies \text{Area}_{\text{target}} = (0.293)^2 \approx 0.086\text{ mm}^2$$

• Step 3: Analyze Clinical Effect: Reducing DFOV from $400\text{ mm}$ to $150\text{ mm}$ decreases pixel linear dimension by $62.5\%$ and pixel area by $86\%$. This reduces pixel sampling size and may reveal detail already supported by the measurements and kernel. It cannot create finer acquired resolution. Noise behavior depends on reconstruction and correlations; changing DFOV alone is not a universal inverse-pixel photon-count law.

---