深度学习重建改善胸部低剂量CT图像质量的价值

Application value of deep learning reconstruction to improve image quality of low-dose chest CT

目的:探讨深度学习重建(DLR)较混合迭代重建(Hybrid IR)在改善胸部低剂量CT(LDCT)图像质量方面的效果。方法:回顾性分析2020年10月至2021年3月在北京协和医院行胸部LDCT体检或因肺内结节定期复查的77例患者。对所有入组患者的影像资料进行不同算法重建,获得标准级别Hybrid IR图像、标准和强级别DLR图像。在3种图像的肺实质、主动脉、肩胛下肌及腋下脂肪内选取感兴趣区并测量其CT值和标准差,用于计算信噪比(SNR)和对比噪声比(CNR)。同时,由2名影像医师按照Likert 5分量表法对图像质量进行主观评分,且记录肺磨玻璃结节(GGN)的数量,并对其显示情况进行评分。2名医师评分不一致时由第3名医师评分决定。采用Kruskal-Wallis非参数检验对3种图像的主观和客观评分进行分析,若总体存在差异,则用Bonferroni校正检验进行组内两两比较。结果:3种图像在肺实质、主动脉、肩胛下肌及腋下脂肪处的CT值差异均无统计学意义( P均>0.05),而图像噪声、SNR和图像的CNR差异均有统计学意义( P均<0.05)。其中标准级别Hybrid IR图像、标准和强级别DLR图像的CNR分别为0.71(0.49,0.88)、1.06(0.78,1.32)和1.14(0.84,1.48)。标准级别和强级别DLR图像均较标准级别Hybrid IR图像的主观和客观噪声低及SNR和CNR高,差异均有统计学意义( P均<0.05)。在对主要解剖结构(肺裂、肺血管、气管和支气管、淋巴结、胸膜和心包)和GGN的显示上,标准级别和强级别DLR图像评分明显优于Hybrid IR图像,差异均有统计学意义( P均<0.05)。 结论:与Hybrid IR相比,DLR可以明显降低LDCT图像的噪声,且对GGN的显示良好,有助于在较低辐射剂量水平时保证图像质量,从而改善采用CT行肺癌筛查及肺结节随访的安全性。

Objective To evaluate the effectiveness of deep learning reconstruction(DLR)compared with hybrid iterative reconstruction(Hybrid IR)in improving the image quality in chest low-dose CT(LDCT).Methods Seventy-seven patients who underwent LDCT scan for physical examination or regular follow-up in Peking Union Medical College Hospital from October 2020 to March 2021 were retrospectively included.The LDCT images were reconstructed with Hybrid IR at standard level(Hybrid IR Stand)and DLR at standard and strong level(DLR Stand and DLR Strong).Regions of interest were placed on pulmonary lobe,aorta,subscapularis muscle and axillary fat to measure the CT value and image noise.The signal to noise ratio(SNR)and contrast to noise ratio(CNR)were calculated.Subjective image quality was evaluated using Likert 5-score method by two experienced radiologists.The number and features of ground-glass nodule(GGN)were also assessed.If the scores of the two radiologists were inconsistent,the score was determined by the third radiologist.The objective and subjective image evaluation were compared using the Kruskal-Wallis test,and the Bonferroni test was used for multiple comparisons within the group.Results Among Hybrid IR Stand,DLR Stand and DLR Strong images,the CT value of pulmonary lobe,aorta,subscapularis muscle and axillary fat had no significant differences(all P>0.05),but the image noise and SNR of pulmonary lobe,aorta,subscapularis muscle and axillary fat had significant differences(all P<0.05),and the CNR of images had significant difference(P<0.05),too.The CNR of Hybrid IR Stand images,DLR stand images and DLR strong images were 0.71(0.49,0.88),1.06(0.78,1.32)and 1.14(0.84,1.48),respectively.Compared with Hybrid IR images,DLR images had lower objective and subjective image noise,higher SNR and CNR(all P<0.05).The scores of DLR images were superior to Hybrid IR images in identifying lung fissures,pulmonary vessels,trachea and bronchi,lymph nodes,pleura,pericardium and GGN(all P<0.05).Conclusions DLR significantly reduced the image noise,and DLR images were superior to Hybrid IR images in identifying GGN in chest LDCT while maintaining superior image quality at relatively low radiation dose levels.Thus DLR images can improve the safety of lung cancer screening and pulmonary nodule follow-up by CT.