Author: Cem Altunbas, Farhang Bayat 👨🔬
Affiliation: Department of Radiation Oncology, University of Colorado School of Medicine, Taussig Cancer Center, Cleveland Clinic 🌍
Purpose:
Low dose imaging and scatter rejection hardware in CBCT reduces X-ray fluence incident on the imager, increasing noise, causing photon starvation artifacts in CBCT images. In this work, we introduced a Non-local Means filtering-based method, Least Filtered yet Denoised Enough Non-Local Means (LFDE NLM), optimized to reduce noise under low X-ray fluence conditions and minimize streak artifacts in CBCT images.
Methods:
A fixed NLM filter weight may cause over- or under- smoothing due to noise variations in projections. In the proposed method, multiple NLM filtered image blocks were generated using a coarse and fine range of NLM filter weights. The optimal filter weight was selected on a pixel-by-pixel basis, based on the relative change in noise magnitude by a given filter weight. To evaluate performance, phantoms and clinical trial participants were CBCT scanned with a high-performance antiscatter grid. Projections were denoised using LFDE NLM followed by filtered backprojection (FBP) reconstruction. Images were also reconstructed using linear filtering (LF) and FBP, and iterative reconstruction (IR). Spatial noise heterogeneity (NH), signal to noise ratio (SNR), and spatial resolution were evaluated.
Results:
Relative to iltered images, LF kept NH the same, IR increased NH by 60±30% with IR, LFDE NLM reduced NH by 30%±10%. Width of the image gradient at the bone-soft tissue interface was 2.1±0.7 mm without filtering, 2.4±0.7 mm for LF, 1.8±0.6 mm for IR, and 2.1±0.5 mm for LFDE NLM. SNR was improved by a factor of 2.9±2.2 for LF and 1.7±2.2 for IR and a factor of 3.7±2.7 for LFDE NLM in regions most affected by photon starvation.
Conclusion:
LFDE NLM significantly reduced photon starvation noise in low X-ray fluence conditions while preserving spatial resolution. It successfully addressed over- and under- denoising challenges with NLM filtering and allowed robust denoising when using FBP based CBCT reconstruction.