Application of deep learning-based image reconstruction methods combine with the asynchronous calibration quantitative CT for the measurement of bone mineral density / 实用放射学杂志

ABSTRACT

Objective To investigate the accuracy and reproducibility of deep learning algorithms combined with asynchronous calibration quantitative computed tomography(QCT)for measuring bone mineral density(BMD),and to explore the feasibility of using low-dose scanning BMD measurement.Methods European spine phantom(ESP)was scanned with asynchronous calibration QCT and conventional synchronous calibration QCT,respectively,the accuracy and short-term reproducibility was compared.ESP were scanned with asynchronous calibration QCT,matching 120 kVp with five sets of tube currents:20,60,100,140,and 180 mA.Three levels of deep learning image reconstruction(DLIR)and hybrid model-based adaptive statistical iterative reconstruction V(40％ASIR-V)were used for reconstruction.The BMD values of three vertebrae in the ESP were measured.Furthermore,the image noise and contrast-to-noise ratio(CNR)were compared.Results The relative errors(RE)of the three vertebrae of the asynchronous calibration QCT and synchronous calibration QCT were all less than 7％.There was no statistical difference in the BMD values of the two scans at one week interval of the asynchronous calibration QCT(P＞0.05).There were no significant differences in RE among different tube currents or different reconstruction methods(P＞0.05).The image quality of deep learning-based image reconstruction of high strength(DLIR-H)at 20 mA tube current was better than that of 40％ASIR-V at 180 mA,and the radiation dose was reduced by 89％.Conclusion Asynchronous calibration QCT has high accuracy in BMD measurement,and has good repeatability.Asynchronous calibration QCT which combined with DLIR does not affect the accuracy of BMD measurement,and can significantly improve the CNR of images and reduce image noise.