Abstract
Filtered backprojection (FBP), one of the most widely used reconstruction methods in tomography, requires a large number of low-noise projections to yield accurate reconstructions. In many applications, complete high-quality projection data cannot be obtained due to practical constraints. While algebraic methods handle such problems better, they are computationally expensive. This paper introduces Minimum Residual Filtered Backprojection (MR-FBP), a method that improves FBP by using a custom data-dependent filter that minimizes the projection error. Experiments show that MR-FBP produces reconstructions with accuracy similar to algebraic methods but at a significantly lower computational cost, effectively bridging the gap between analytical and iterative reconstruction techniques.
Introduction: The Tomography Trade-off
In tomographic imaging, such as CT scans or electron microscopy, a 3D object is reconstructed from a series of 2D projections. Traditionally, practitioners face a choice between two distinct approaches:
- Analytical Methods (FBP): These are fast and computationally efficient but rely on static filters (like the Ram-Lak ramp filter). They perform poorly when data is "sparse" (few angles) or noisy, resulting in significant streak artifacts.
- Algebraic Methods (e.g., SIRT): These treat reconstruction as an iterative optimization problem. They are robust to noise and limited data but are computationally intensive, often taking orders of magnitude longer than FBP.
The authors propose Minimum Residual Filtered Backprojection (MR-FBP) to achieve the robustness of algebraic methods with the speed of FBP.
Methodology: Data-Dependent Filtering
The core innovation of MR-FBP is the calculation of an optimal filter ($h^*$) tailored to the specific dataset being reconstructed. Instead of a "one-size-fits-all" static filter, MR-FBP finds a filter that minimizes the "projection error"—the difference between the measured data and the projections that would be generated by the resulting reconstruction.
Because FBP is a linear operation, the authors demonstrate that finding this optimal filter can be reduced to solving a small linear system. To ensure the method remains fast, they employ exponential binning.
By recognizing that the most critical information in a tomographic filter is concentrated near its center, the authors use small bins at the center and exponentially larger bins further out. This reduces the number of unknowns in the filter calculation to $O(\log N_d)$, making the optimization problem trivial to solve.
Key Results and Performance
The method was evaluated using simulated phantoms (like the Shepp-Logan head) and real-world $\mu$-CT data from mouse femurs and diamonds.
Key findings include:
- Speed: MR-FBP was approximately 20 times faster than the SIRT algebraic method while delivering comparable image quality.
- Robustness: In tests with as few as 32 projections, MR-FBP significantly outperformed standard filters (Ram-Lak, Hann) in terms of Mean Absolute Error and Structural Similarity Index (SSIM).
- Simplicity: Unlike iterative methods that require tuning the number of iterations, basic MR-FBP is essentially parameter-free.
Conclusion
MR-FBP represents a significant advancement in tomographic reconstruction, offering a high-performance alternative to both standard FBP and slow iterative solvers. By making the filter data-dependent and using efficient basis functions like exponential binning, the method provides a practical solution for high-quality imaging under constrained data conditions. Future work may extend this approach to 3D cone-beam geometries and explore alternative basis functions for specialized applications.
Summary updated