Preliminary Study Of Ultralow-Dose CT Based On Cycle-Consistent Generative Adversarial Network For Pulmonary Tuberculosis Denoising

Objective:Tuberculosis is a highly infectious disease with a high mortality rate.At present,CT examination is a routine examination for the diagnosis and follow-up.Since multiple CT examinations are prone to increase the radiation dose,how to reduce the radiation dose of CT scanning while ensuring the image quality has been the long-term goal of imaging research.With the development of artificial intelligence,deep learning has been applied in various fields of medical image.The purpose of this study was to develop a deep learning method based on the Cyclegenerative adversarial network(Cycle-GAN)to denoise ultra-low dose CT images for pulmonary tuberculosis.Materials and Methods:CT images were prospectively obtained from 129 patients with pulmonary tuberculosis(Age:42.6±13.9 years,BMI:21.5±4.5 kg/m2,Males:74,Females:55)who underwent chest CT examination from 2019 to 2022.Two consecutive CT scans were performed at different doses(normal dose:tube voltage of 120kVp with automated exposure control;ultra-low dose:tube voltage of 80kVp and tube current of 10 mAs).The scanned CT images were randomly divided into two groups.The training group was used to train the Cycle-GAN network to synthesize images,and the validation group was used to test the performance of the network.The Cycle-GAN synthesized images were compared with the hybrid iterative reconstruction algorithm,model-based iterative reconstruction,and filtered back projection method to compare objective measurement and subjective score.Repeated Measures ANOVA was used for objective measurements,and Friedman test was used for comparison of subjective scores,Objective data measures(noise,signal-to-noise ratio,contrast-to-noise ratio,figure of merit,peak signal to noise ratio,structural similarity of aorta,lesion,airway,muscle,fat)and subjective scores(overall image quality,noise,artifact,blurring effect,small vessels,lymph nodes,pericardium,pulmonary fissure,pulmonary nodules,fibrosis,lesions,calcification,lymph node enlargement,ground glass nodule)were compared.Results:When Cycle-GAN network was used for image denoising,the mean peak signal-to-noise ratio(PSNR)was 32.8±3.4 and the mean structural similarity index(SSIM)was 0.797±0.05,both of which were significantly higher than the other three algorithms(PSNR,IMR:32.2±3.2,iDose:31.3±2.8,FBP:28.2±2.2,p<0.001;SSIM,IMR:0.792±0.05,p=0.21,iDose:0.627±0.06,FBP:0.435±0.08,p<0.001).The noise level(38.5±6.1 HU)of the Cycle-GAN network was lower than that of the other three algorithms(IMR:47.7±11.6,iDose:83.5±12.7,FBP:160.4±42.7,p<0.001).The SNR(1.19±0.4)was higher than that of the other three algorithms(IMR:1.02±0.3,iDose:0.64±0.2,FBP:0.39±0.1,all p<0.001).The total dose of the two scans was within the national standard chest CT dose range and ultralow-dose CT scans showed a 96%reduction in effective dose(ED)compared with normal dose CT scans(EDULD:0.1±0.0 mSv,EDND:2.3±0.7 mSv,p<0.001).Conclusion:The image quality of ultra-low dose CT images based on Cycle-Gan network is better than IMR,iDose and FBP reconstruction methods for pulmonary tuberculosis,which can satisfy the diagnosis of tuberculosis patients by radiologists.