Python Code: CT-Reconstruction-Benchmark

Python Code for the following research work is available athttps://github.com/code-depository/CT-Reconstruction-BenchmarkDescription of the work:Title: Filter selection versus iterative reconstruction in CT: a noise-dose benchmark using open-source Python and clinical DICOM validationExecutive Summary:- Background: Filtered backprojection (FBP) remains the workhorse of clinical CT reconstruction, yet no systematic open-source benchmark has compared all major FBP filter variants against iterative reconstruction across a clinically calibrated dose ladder.Objective: To determine the dose threshold D* at which SART achieves a structurally meaningful advantage over the best FBP filter, and to identify which FBP filter most closely tracks SART across the dose ladder.Methods: A fully reproducible Python pipeline was developed using scikit-image and ASTRA Toolbox. The study comprised three experimental arms. Arm 1 (phantom, parallel-beam): five algorithms (IRT, FBP/ramp, FBP/Shepp-Logan, FBP/cosine, SART) were evaluated on a 400x400 Shepp-Logan phantom at six simulated dose levels (100% to 1%, I0 = 5x10^4 to 200 photons per detector element) with five independent Poisson noise realisations per cell (150 total reconstructions). Arm 2 (DICOM validation): two axial brain CT slices from the Kaggle brain stroke dataset (one normal, one intracerebral haemorrhage) were reconstructed at full dose with HU accuracy confirmed across four tissue ROIs in each case. Arm 3 (fan-beam benchmark): SART was repeated under fan-flat geometry using ASTRA Toolbox (SID = 570 px, SDD = 1040 px, 360 angles) to confirm geometry independence of the ranking. The primary metric was SSIM (data_range = 1.0).Results: FBP/cosine, not FBP/Shepp-Logan, was the superior analytic filter at all dose levels (SSIM 0.761 vs 0.580 at full dose). SART significantly outperformed FBP/cosine at every dose level tested under both parallel-beam and fan-beam geometry, with delta-SSIM = 0.134 (parallel) and 0.309 (fan-beam) at full dose. D* = 100% under both geometries, indicating SART is superior across the entire clinical dose range. DICOM validation achieved 32/32 HU checks passed across both cases; FBP/cosine achieved SSIM of 0.990 (normal) and 0.975 (ICH) against the original clinical slices.Conclusion: FBP/cosine is the optimal analytic filter on the SSIM metric, and SART demonstrates significant structural fidelity advantages at all dose levels. The complete pipeline, data, and figures are openly available for reproducibility.Keywords: CT reconstruction; filtered backprojection; iterative reconstruction; SART; SSIM; low-dose CT; open-source; Python; scikit-image