Feasibility Study of Deep Learning Reconstruction To Improve Dual-energy CT Image Quality and Lesion Diagnosis
Objective To evaluate the feasibility of deep learning image reconstruction(DLIR)on improving image quality and lesion diagnosis using virtual monochromatic spectral images in abdominal dual-energy CT(DECT),compared with adaptive statistical iterative reconstruction-V(ASIR-V).Methods Sixty-five patients who completed abdominal dual-energy CT scan were randomly included and reconstructed by ASIR-V40％,DLIR-M(moderate),and DLIR-H(height).The portal images of conventional 5mm ASIR-V40％at 70 keV and ASIR-V40％,DLIR-M and DLIR-H of thinner layer(1.25mm)at 70 keV.Measure CT attenuation,standard deviation(SD)value,signal-to-noise ratio(SNR),and noise to contrast(CNR)of liver,spleen,vertical spine,and intramuscular fat.The number of liver lesions in the portal stage images of thinner layer groups was counted.Image quality and diagnosis confidence were subjectively evaluated by two radiologists with extensive experience.Results For the 1.25mm images with 70keV,DLIR-M and DLIR-H had lower SD,higher SNR and CNR,and better subjective image quality than ASIR-V40％with consistent lesion detection rates and DLIR-H performed the best(all P＜0.001).There was no significant statistical difference in the SD value,CNR between 5mm ASIR-V40％group at 70 keV and 1.25mm DLIR-H group at 70 keV(P=0.211,0.358,0.208,0.052).Conclusion Compared with the conventional ASIR-V,deep learning reconstruction algorithm(DLIR)with DECT can further reduce the image noise of abdominal CT obtain better image quality and higher confidence in lesion diagnosis.Moreover,DLIR-H at 70 keV can achieve thinner thickness image reconstruction with similar image noise compared with ASIR-V40％at 70 keV.
Deep Learning Image ReconstructionSpectral ImagingPortal PhaseImage QualityLesions Diagnosis
万方数据
维普
