基于深度学习重建算法对上腹部CT图像质量的研究

Study of the Quality of Upper Abdominal CT Image Reconstructed by Deep Learning

目的:对比常规新一代自适应迭代重建算法(ASiR-V),研究基于深度学习的图像重建算法(DLIR)对上腹部增强CT图像质量和诊断信心的提高。方法:纳入30例行上腹部CT增强的患者,对其门静脉期原始数据使用滤波反投影(FBP),30%ASiR-V,70%ASiR-V和DLIR-低(L)、中(M)、高(H)3个重建等级分别进行重建。测量肝脏、脾脏和右侧竖脊肌的CT值,计算相应背景噪声(SD)值、信噪比(SNR)值和对比噪声比(CNR)值。2名放射科医师分别评价6组重建图像的图像噪声、细小结构显示及整体感观。结果:(1)客观评价:6组重建图像在肝脏、脾脏及右侧竖脊肌的SD值、CNR值和SNR值的差异均有统计学意义(P<0.001)。DLIR-H最优,DLIR-M与70%ASiR-V无显著性差异。随DLIR重建强度增加,SD值降低,SNR值和CNR值升高。(2)主观评价:2位放射科医师对图像质量的主观评价一致性尚可(kappa值分别为0.54、0.59、0.62)。DLIR-M和DLIR-H表现出最佳主观图像质量分数。结论:DLIR与FBP、ASiR-V算法相比,降噪效果更好,图像质量更高。DLIR-M和DLIR-H的重建图像提供了最优的主客观图像质量数据。

Purpose: To investigate the deep learning image reconstruction(DLIR) on image quality and diagnostic confidence of upper abdominal enhanced CT compared with adaptive statistical iterative reconstructionveo(ASiR-V). Methods: Thirty patients underwent upper abdominal CT enhancement were selected. Raw data of portal vein stage were respectively reconstructed with filtered back projection(FBP), 30% ASiR-V, 70% ASiR-V and DLIR(L, M and H) techniques. standard deviation(SD), signal to noise ratio(SNR), and contrast to noise ratio(CNR) values were obtained from liver, spleen, and right erector spinal muscle at three consecutive levels, which showed the best upper abdominal structure. The 5-point method was used to evaluate the image noise, the display of small structures and the overall perception.Results: The differences of SD, CNR and SNR in liver, spleen and right erector spinae were with statistical significant among the six reconstructed images(P<0.001). DLIR-H was the best.There was no significant difference between DLIR-M and 70% ASiR-V. As the strength levels of DLIR increased,SD decreased, SNR and CNR increased. The evaluation of image quality by two radiologists was consistent(kappa=0.54, 0.59, 0.62). DLIR-M and DLIR-H demonstrated the best subjective image quality scores. Conclusion:Compared with FBP and partial ASi R-V, DLIR is with better noise reduction effect and higher image quality. The reconstruction images of DLIR-M and DLIR-H provide the best subjective and objective image quality.