基于深度学习的ClearInfinity算法在肝脏病变低剂量CT扫描中的应用价值

Study on Application Value of ClearInfinity Algorithm Based on Deep Learning in Low-dose CT Scanning of Liver Lesions

目的 探究基于深度学习的ClearInfinity(CI)算法在肝脏病变低剂量CT扫描中改善图像质量的价值。方法 前瞻性收集在我院行上腹部增强CT检查的肝脏病变患者65例,分别在动脉期(Arterial Phase,AP)和门脉期(Portal Phase,PP)采用常规剂量,延迟期(Delayed Phase,DP)采用低剂量进行腹部三期增强扫描。根据重建方法的不同,将动脉期常规剂量组(A组)采用迭代算法(50%ClearView,50%CV)重建;延迟期低剂量组(B组)分为B1、B2两个亚组,B1组采用迭代算法(50%CV)重建,B2组采用深度学习算法(50%CI)重建。比较A、B两组辐射剂量的差异。测量B1组、B2组肝实质、脾脏及肾脏的CT值及A组、B1组、B2组3组腹部皮下脂肪噪声值,计算各感兴趣区(Region of Interest,ROI)信噪比(Singal to Noise Ratio,SNR)和对比噪声比(Contrast to Noise Ratio,CNR)。结果 与A组比较,B组辐射剂量降低约73.3%,差异具有统计学意义(P<0.05)。各组间噪声比较:B1组噪声最大,B2组<A组<B1组,但A组与B2组间差异无统计学意义(P=0.625);B1组、B2组间肝脏、胰腺及肾脏CT值基本保持不变,差异无统计学意义(P>0.05)。SNR值和CNR值比较,B2组各ROI SNR和CNR值均高于B1组(P<0.001)。3组间图像主观评分差异有统计学意义(P<0.001),B1组各评分均<3分,无法满足诊断需要;而B2组与A组间各主观评分差异无统计学意义(P>0.05)。结论 在辐射剂量大幅降低的情况下,采用基于深度学习的50%CI算法可以显著降低肝脏病变低剂量CT扫描时的图像噪声,改善图像质量,满足诊断需要。

Objective To explore the application value of ClearInfinity(CI)algorithm based on deep learning to improve image quality in low-dose CT scanning of liver lesions.Methods 65 patients with hepatic disease who underwent contrast-enhanced liver CT in our hospital were retrospectively collected.Conventional dose protocol was used in arterial phase(AP)and portal phase(PP),and low dose protocol of abdominal three-phase enhanced scanning was used in delayed phase(DP).According to the different reconstruction methods,iterative algorithm(50%ClearView,50%CV)was used to reconstruct the conventional dose group(group A)in arterial phase.The delayed low-dose group(group B)was divided into two subgroups:group B1 and group B2.Group B1 was reconstructed by iterative algorithm(50%CV),and group B2 was reconstructed by deep learning algorithm(50%CI).The difference of radiation dose of group A and group B were recorded and compared.The CT values of liver parenchyma,spleen and kidney in groups B1 and B2 and the subcutaneous fat noise in groups A and B were measured,and the signal to noise ratio(SNR)and the contrast to noise ratio(CNR)of all region of interest(ROI)were calculated.Results Compared with group A,the radiation dose of group B was reduced by 73.3%,and the difference was statisticallysignificant(P<0.05).Comparison of noise among groups:the noise of group B1 was the highest,and that of group B2<group A<group B1,but there was no statistically significant differences between group A and B2(P=0.625).CT values of liver,pancreas and kidney were basically stable between groups B1 and B2,with no statistically significant differences(P>0.05).The SNR and CNR of all ROIs in group B2 were all higher than those of group B1(P<0.001).Subjective image scores of difference groups were statistically different(P<0.001).The scores of group B1were<3 scores,which could not meet the needs of diagnosis.While there were no significant differences in the subjective scores between group B2 and group A(P>0.05).Conclusion When the radiation dose was greatly reduced,the 50%CI algorithm based on deep learning could reduce the image noise of low-dose CT liver scanning,improve the image quality and satisfy diagnostic requirement.