深度学习算法在超重肺间质性病变患者中的应用研究

Application of Deep Learning Algorithm in Overweight Patients with Interstitial Lung Disease

目的探索深度学习算法提高超重肺间质性病变患者的低剂量CT(Low Dose CT,LDCT)及高分辨率CT(High Resolution CT,HRCT)扫描图像质量的应用效果。方法前瞻性地纳入20例超重肺间质病(Interstitial Lung Disease,ILD)病例,其中5例为结缔组织病相关ILD,余15例病因不明。所有患者均接受HRCT扫描(120 kVp,自动管电流)和LDCT扫描(120 kVp,30 mAs)。HRCT扫描图像由混合迭代重建算法(Adaptive Iterative Dose Reduction 3-Dimensional,AIDR3D)处理,LDCT图像由深度学习重建算法(Advanced Intelligence Clear-IQ Engine,AiCE,肺/骨算法,mild/standard/strong重建)处理。两名放射科医师分别对图像的噪声、伪影、图像质量、正常结构以及ILD相关的特征表现进行评估,比较LDCT扫描组与HRCT扫描组的图像噪声与图像质量。结果超重ILD患者应用LDCT扫描方案后,有效扫描剂量比HRCT扫描下降73%。其中,LDCT扫描中,应用肺算法(standard/strong重建)以及骨算法(mild/standard/strong重建)处理的图像噪声约降至HRCT扫描的34.5%~91.7%(P<0.05),信噪比(范围:24.64~66.23)约为后者(平均值22.75)的1.1~2.9倍(P<0.001)。两组扫描方案的重建图像主观评分(图像总体评分、伪影、正常结构观察如叶间裂、近端支气管及邻近肺血管、外周支气管及邻近肺血管、胸膜下血管等)均未表现出显著差异(P>0.05)。在异常征象的评估中,LDCT扫描(肺算法,strong重建)对磨玻璃影的观察显著优于HRCT扫描(P=0.002),在其他异常特征(网格影、支气管扩张及蜂窝征)的观察中,两种方案的图像质量没有显著差异。结论在超重ILD患者中,应用深度学习算法可以在有效降低辐射剂量的同时,保证图像质量。

Objective To explore the application effect of deep learning algorithm in improving the image quality of low-dose CT(LDCT)scans and high resolution CT(HRCT)scans in patients with overweight interstitial lung disease(ILD).Methods A total of 20 cases of overweight ILD patients were prospectively enrolled in this study,in which 5 patients were connective tissue disease-related ILD,the remaining 15 cases of unknown etiology.All patients underwent HRCT(120 kVp,automatic tube current)followed by LDCT(120 kVp,30 mAs).HRCT images were reconstructed with adaptive iterative dose reduction 3-dimensional(AIDR3D).and LDCT images were reconstructed with advanced intelligence clear-IQ engine(AiCE lung/bone algorithm,mild/standard/strong).The image noise,streak artifact,image quality,visualization of normal and abdominal features of ILD were evaluated by two radiologists,and the image noise and image quality were compared between LDCT and HRCT scans groups.Results Patients with overweight ILD received a 73%reduction in the effective dose of LDCT scans compared with HRCT scans.The image noise of LDCT(lung-standard/strong,bone-mild/standard/strong)was reduced to 34.5%to 91.7%of the HRCT(P<0.05),and signal to noise ratio(range:24.64-66.23 vs.22.75)was about 1.1 to 2.9 times increased(P<0.001).The overall image quality,streak artifact,and visualization of normal features(fissures,proximal bronchi and vessels,peripheral bronchi and vessels,subpleural vessels)between all reconstructed LDCT images and HRCT images showed no significant difference(P>0.05).LDCT(AiCE,lung-strong)was superior to HRCT(AIDR3D)in visualizing ground glass opacity(P=0.002),while other abnormal structures(reticulation,bronchiolectasis and honeycombing)were similar between reconstructed LDCT and HRCT images.Conclusion In overweight ILD patients,deep learning reconstruction can be applied to reduce the radiation dose significantly and keep the image quality.