摘要目的:探究深度学习算法（DLR）在胰腺低剂量CT扫描中改善图像质量的价值。方法:前瞻性收集2020年8月至12月在北京协和医院进行腹部胰腺增强检查的68例患者，采用随机区组法分为正常剂量组和低剂量组，各34例，正常剂量组扫描管电压为120 kV，低剂量组管电压为100 kV。所有患者均行平扫、动脉期、实质期、延迟期扫描。低剂量组4期图像分别采用滤波反投影（FBP）、混合模型迭代算法（AIDR）和DLR 3种方法重建图像，正常剂量组4期图像均采用AIDR重建图像，分别记为LD-FBP、LD-AIDR、LD-DLR、RD-AIDR。测量胰腺的CT值、噪声值（SD），计算信噪比（SNR）、对比噪声比（CNR），不同重建方法图像间各参数的比较采用单因素方差分析，两两比较采用LSD法；对每组图像进行主观评分，多组间比较采用Kruskal-Wallis检验。结果:正常剂量组和低剂量组患者胰腺平扫、动脉期、实质期、延迟期不同重建方法图像间胰腺CT值、SD、SNR、CNR差异均有统计学意义（ P均<0.05）。实质期和延迟期图像LD-FBP、LD-AIDR、LD-DLR图像CT值均高于RD-AIDR图像（ P均<0.05）；4期图像间SD、SNR两两比较差异均有统计学意义（ P均<0.05）；4期LD-FBP、LD-DLR与RD-AIDR的CNR差异有统计学意义（ P均<0.05），RD-AIDR的CNR优于LD-FBP，LD-DLR优于RD-AIDR。DLR算法对胰腺4期图像的SD、SNR、CNR均有改善，其中对SNR的改善增强图像更为显著，对CNR的改善平扫期和延迟期更为显著。胰腺4期不同算法重建图像主观评分差异均有统计学意义（ P均<0.001）。LD-DLR与RD-AIDR评分差异无统计学意义（平扫、动脉期、实质期、延迟期 Z值分别为1.00、2.24、0.45、1.34， P值分别为0.317、0.025、0.655、0.180）。 结论:DLR技术可以在降低胰腺CT辐射剂量的同时改善图像质量，满足诊断需求，在胰腺低剂量增强CT检查中可降低噪声，提高图像的信号SNR和CNR。

abstractsObjective:To explore application value of improving quality of the low dose pancreatic CT images by using deep learning reconstruction (DLR).Methods:From August to December 2020, 68 patients who underwent contrast-enhanced pancreatic CT were prospectively collected in Peking Union Medical College Hospital. All patients were randomly divided into routine dose group (34 patients, with tube voltage of 120 kV) and low dose group (34 patients, with tube voltage of 100 kV). All patients underwent non-contrast, arterial phase, parenchymal phase and delay phase scans. The four-phase images of low dose group were reconstructed by using filtered back projection (FBP), hybrid iterative reconstruction (AIDR) and DLR which were marked with LD-FBP, LD-AIDR and LD-DLR, respectively. The four-phase images of routine dose group were reconstructed by using AIDR algorithm which were marked with RD-AIDR. The CT value, image noise (SD), signal to noise ratio (SNR) and contrast to noise ratio (CNR) of pancreas were measured. The ANOVA test was performed in comparison with objective parameters of different reconstruction algorithms, and LSD test was performed in pairwise comparison. The subjective image scores were obtained and were compared using Kruskal-Wallis test.Results:CT value, SD, SNR and CNR of non-contrast, arterial phase, parenchymal phase and delay phase had significant difference among different reconstruction images of routine dose group and low dose group (all P<0.05). The CT value of LD-FBP, LD-AIDR, and LD-DLR images were significantly higher than those of RD-AIDR images in parenchymal phase and delay phase (all P<0.05). There were statistically significant differences in each pairwise comparison of SD and SNR of four phase images (all P<0.05). There were statistically significant differences of CNR among LD-FBP, LD-DLR and RD-AIDR in four phase images (all P<0.05). The CNR of RD-AIDR was better than that of LD-FBP, and CNR of LD-DLR was better than that of RD-AIDR. DLR algorithm improved the SD, SNR and CNR of four phases of pancreatic images. The improvement of SNR was more significant after contrast enhancement, and the improvement of CNR was more significant in the non-contrast and delay phases. Subjective image scores of different reconstruction images were statistically different in four phase images (all P<0.001). Overall image scores of LD-DLR and RD-AIDR had no significant differences in four phase ( Z value of four phases were 1.00, 2.24, 0.45 and 1.34, respectively; P value of four phases were 0.317, 0.025, 0.655 and 0.180, respectively). Conclusion:The DLR technology can decrease radiation dose of pancreatic CT, improve image quality and satisfy diagnostic requirement. The DLR technology can also reduce image noise, improve the SNR and CNR in low dose contrast-enhanced pancreatic CT.