A deep-learning method for the denoising of ultra-low dose chest CT in coronary artery calcium score evaluation

Abstract

To evaluate a novel deep-learning denoising method for ultra-low dose CT (ULDCT) in the assessment of coronary artery calcium score (CACS). Sixty adult patients who underwent two unenhanced chest CT examinations, a normal dose CT (NDCT) and an ULDCT, were enrolled prospectively between September 2017 to December 201. A special training set was created to learn the characteristics of the real noise affecting the ULDCT implementing a fully convolutional neural network with batch normalisation. Subsequently, the 60 ULDCTs of the evaluation set were denoised. Two blinded radiologists assessed the NDCT, ULDCT, and denoised-ULDCT (DULDCT), assigning a CACS and categorised each scan as having a score above or below 100 and presence of calcifications (score 0 versus >0). Statistical analysis was used to evaluate the agreement between the readers and differences in CACSs between each imaging method. After excluding one patient, the cohort included 59 patients (median age 67 years, 58% men). The ULDCT median effective radiation dose (ERD) was 0.172 mSv, which was 2.8% of the NDCT median ERD. Denoising improved the signal-to-noise ratio by 27.7% (p<0.001). Interobserver agreement was almost perfect between readers (intraclass correlation coefficient >0.993). CACSs were lower for ULDCT and DULDCT as compared to the NDCT (p≤0.001). In differentiating between the presence and absence of coronary artery calcifications, DULDCT showed greater accuracy (98-100%) and positive likelihood ratio (14.29->99) compared to ULDCT (92% and 2.78, respectively). DULCT significantly reduced the image noise and better identified patients with no coronary artery calcifications than native ULDCT.