Room: Davidson Ballroom B
Purpose: Quality enhancement is important to push the ALARA boundary in low-dose CT quantitative imaging (QI). We previously developed a modified block-matching 3D (BM3D) algorithm and demonstrated superior SNR performance in low-dose CT denoising. This study investigates its clinical implication in QI context by applying the modified BM3D method to emphysema scoring in thoracic CT screening, and comparing it with the basic BM3D scheme.
Methods: We evaluated the efficacy of the low-dose CT denoising method on emphysema scoring over 20 cases, among which 10 cases had pronounced emphysema. For each case, volumetric image datasets with both the original (2 mGy) dose and the simulated dose reduced to only 5% of the original (0.1 mGy) were used as input. The low-dose CT images were generated using a verified pipeline by adding realistic stochastic noise to the raw data in the projection domain. The modified BM3D denoising method was applied to the low-dose images, and the resulting emphysema scoring was compared against that from the full-dose images and that from denoising using the basic BM3D scheme.
Results: Visual inspection of the emphysema mask showed that quality degradation caused by low-dose imaging resulted in a significant portion of voxels being mistakenly labeled as emphysema. The modified BM3D method effectively corrected voxels without emphysema while giving a decent detection on the abnormal area. The quantitative emphysema scoring further demonstrated that the modified BM3D method reduced the scoring deviation from full-dose imaging by 58%. It had statistically significant advantage over that using the basic BM3D scheme, yielding a p-value of 6.30E-5 with the well-investigated measure of RA950.
Conclusion: The modified BM3D denoising algorithm brings significant benefit in emphysema scoring on ultra-low-dose CT images. Its rationale is generally applicable to other quantitative imaging tasks, where images are corrupted by non-Gaussian noise.
Low-dose CT, Quantitative Imaging, Image Processing