FreeCT_wFBP

FreeCT_wFBP implements weighted filtered backprojection (wFBP) to reconstruct helical fan-beam CT data for quantitative imaging, radiomics, and computer-aided detection.

Key Features:

• Helical scan support: Reconstructs helical scans with arbitrary pitch and supports flying focal spots and quarter-detector offset sampling.
• Weighted filtered backprojection (wFBP): Implements wFBP adapted for helical fan-beam CT geometry.
• GPU-based processing: Host code is written in C with GPU kernels implemented in Nvidia CUDA C to accelerate reconstruction for high-resolution datasets.
• Low-dependency architecture: Minimal external software dependencies.
• Reconstruction speed: Achieves an average reconstruction time of 17.4 ± 1.0 seconds for a volume of 512 rows × 512 columns × 32 slices measured using the ACR CT accreditation phantom across four flying focal spot configurations.
• Image quality validation: Meets CT Accreditation Program criteria for CT number accuracy, uniformity, and contrast-to-noise ratio (CNR) based on ACR phantom assessments.
• Mass-attenuation accuracy: FORBILD thorax phantom simulations yielded reconstructed water mass-attenuation coefficient values within 0.0001 mm²/g (0.01%) of known phantom values.

Scientific Applications:

• Quantitative imaging and radiomics: Provides reconstructions suitable for quantitative measurements and radiomics feature extraction from CT images.
• Computer-aided detection (CAD): Supplies accurate reconstructed images for CAD algorithm development and evaluation.
• Accreditation and validation studies: Supports CT Accreditation Program validation and reconstruction accuracy studies using ACR and FORBILD phantoms.
• Acquisition and reconstruction research: Enables investigation of complex third-generation CT acquisition techniques and sampling patterns such as flying focal spots and quarter-detector offset.

Methodology:

Reconstruction uses weighted filtered backprojection (wFBP) adapted for helical fan-beam geometry with support for arbitrary pitch, flying focal spots, and quarter-detector offset; implementation uses C host code and Nvidia CUDA C GPU kernels and was evaluated with ACR CT accreditation phantom and FORBILD thorax phantom benchmarks, including timing for 512 rows × 512 columns × 32 slices.