A Study on Optimizing the Smoothing Factor for Fast Non-local Means-based Denoising Algorithm in Brain CT Phantom Image using 3D Printing

The purpose of this study is to fabricate a 3D printed phantom for computed tomography (CT) imaging and to optimize the smoothing factor of the fast non-local means (FNLM) denoising algorithm. For this purpose, a 3D printer with fused filament fabrication technology was used to produce the brain CT phantom with a diameter of 16 cm that mimics cerebrospinal fluid, gray matter, white matter, and bone. Then, we added the Gaussian noise of 0.005 to express the low-dose conditions from CT images acquired using a brain phantom. The optimization of performed by applying the FNLM algorithm to the noise-added CT phantom image. The d-value for adjusting smoothing intensity of FNLM algorithm was applied from 0.01 to 1.00 in increments of 0.01. The contrast to noise ratio (CNR), signal to noise ratio (SNR), and pixel profiles were measured for quantitative evaluation. As a result, both CNR and SNR were greatly improved until the d-value was 0.07, however, gradual changes were observed showing a flattened slope after 0.07. In addition, we confirmed that the signal value was stabilized at 0.07 in the measurement of the pixel profile. The incline exhibited a milder at subsequent d-values and the contrast was decreased. Based on these results, the smoothing factor of the FNLM algorithm was optimized to 0.07. Additionally, a comparative evaluation with the Wiener filter, median modified Wiener filter and total variation algorithm was conducted. As a result, the most improved CNR and SNR values were obtained in all tissues when applying the FNLM algorithm with an optimized d-value of 0.07. In conclusion, the FNLM algorithm with d-value of 0.07 was verified to be effective in reducing noise for low-dose CT conditions.