ABSTRACT
Objective To assess the ability of a deep machine learning technique for improving the quality of one-stop renal low dose CTP images.Methods Twenty-one cases who underwent renal noncontrast CT,triple-phase contrast enhanced CT,and CT perfusion (CTP) were collected prospectively.Revolution CT scanner was used with the scan protocol as followed:120 kVp,20 mA for CTP and 100 mA for triple-phase conctrast enhancement,axial scan,ASIR-V80％,rotation 0.5 s,coverage area for z-axial 160 mm,thickness 5 mm.A total of 15 phases were obtained for the first 28 s and then scanned once at 39,43,47,51,63,83,113,213,353,593 s for CTP,which the phases at the 22,51 and 153 s were the cortical phase,medullary phase and excretory phase,respectively.All CTP data was reconstructed with a deep machine learning technique pixel shine A7 model.The data before and after reconstruction was in group A and in group B,respectively.Compared the all data of cortex in the cortical phase and CTP parameters between the two groups.Results There were significant differences of CT values of SD of cortex (9.04 ± 1.77 and 5.75 ± 1.00,respectively),CT values of SD of elector spinae (8.52 ±2.28 and 5.67 ±0.98,respectively),CNR(16.28 ±6.61 and 28.90 ±1.50,respectively) and SNR (21.41 ± 6.67 and 30.65 ± 7.67,respectively) between the two groups (t=1.562,6.286,5.925,-5.892,-17.274,P＜0.05).The SD of images after PS-B was lower than that before PS-B significantly and SNR was improved obviously.There were no differences of cortical blood flow (BF),blood volume (BV),time to peak (TP) and medullary permeability of surface (PS) between the two groups (P ＞ 0.05).Conclusions The reconstruction of deep machine learning PixelShine technique PS-A7 can reduce the noise of images obtained with low tube current,improve the SNR and can not effect the CTP parameters.
ABSTRACT
Purpose To explore the feasibility of denoising algorithm-PixelShine algorithm based on deep learning to enhance the quality of abdominal arterial phase CT images rebuilt by 70 kVp combined with adaptive statistical iterative reconstruction-Veo (ASiR-V). Materials and Methods Abdominal arterial phase images of 33 patients [body mass index (BMI) BMI≤20 kg/m2] scanned by GE Revolution CT were retrospectively analyzed (group A) using 70 kVp tube voltage and 50% ASiR-V technique. PixelShine algorithm B2 mode was applied to post-process group A images to obtain PixelShine image (group B). Two observers rated the image quality of the two groups via a 5-point rating system. The consistency of the rating was analyzed. The difference in ratings, noise, virtual signal-to-noise ratio (SNR) of liver and pancreas and contrast noise ratio (CNR) were compared between the two groups of images. Results The image quality rating of group A and B were(3.12±0.33) scores and(3.97±0.53) scores respectively,noise value(14.50±1.42)HU vs(10.05±1.80)HU, liver virtual SNR 4.51±0.53 vs 6.78.±1.27,liver virtual CNR 0.89±0.55 vs 1.42±0.81,pancreatic virtual SNR 9.51±1.69 vs 13.87±3.26, and pancreatic virtual CNR 5.83±1.66 vs 8.48±2.46.The quality rating of images,liver and pancreas virtual SNR,CNR in group B were all higher than those in group A, and the image noise of group B decreased about 31% compared with that of group A, the difference was statistically significant (P<0.05). Conclusion Post-processing with PixelShine algorithm can improve the image quality of 70 kVp abdominal arterial phase, significantly reduce image noise, and increase image SNR and CNR.
ABSTRACT
Objective To investigate the impact of deep machine learning Pixel Shine (PS) algorithm on image quality of abdominal low-dose plain CT scanning in BMI≥25 kg/m2 patients.Methods A total of 59 patients (BMI≥25 kg/m2) who underwent abdominal CT scan were collected.The patients were divided into group A (100 kVp,n=30) and B (120 kVp,n=29) according to the tube voltage.According to different reconstruction algorithms and treatment methods,patients in group A were divided into A1 (FBP),A2 (FBP+PS),A3 (50％ASiR-V) and A4 (50％ASiR-V+PS) subgroups,while in group B were divided into B1 (FBP) and B2 (50％ASiR-V) subgroups.CT and SD values of right hepatic lobe and right erector spinae were measured,then SNR and CNR of liver and CT dose index of volume (CTDIvol) were calculated.The consistency of parameters measured by two observers was evaluated.Results The consistency of parameters measured by two observers was good (all ICC＞0.80).There was no statistical difference of CT values of liver and erector among A1-A4 subgroups (all P＞0.05),whereas statistical differences of SD values of liver and erector spinae,also of SNR and CNR of liver were found (all P＜0.001).Among A1-A4 subgroups,SDA4 ＜SDA2 ＜SDA3 ＜SDA1,SNRA4 ＞SNRA2 ＞SNRA3 ＞ SNRA1 (all P＜0.001) was observed.There was no significant difference of CNR between A1 and A3 subgroup (P=0.078),while CNRA4＞ CNRA2＞ CNRA3 or CNRA1 (P＜0.001) was noticed.SD values of the liver in subgroup A2 was lower than subgroup B1,and A4 was lower than B2 subgroup (all P＜0.001),and SNR and CNR increased significantly in A2 and A4 subgroups (all P＜0.001).CTDIvol of group A was lower than that of group B (P＜0.001).Conclusion Deep machine learning PS algorithm can improve image quality of abdominal low-dose plain CT scanning in high-BMI patients.