| Part 1 The Construction of Deep Learning Image Super-resolution ModelObjectiveTo provide super-high-resolution images of lung without additional CT scan,based on generative adversarial network(GAN)algorithm,a deep learning model was constructed to generate super-resolution CT(SRCT)images from high resolution CT(HRCT).Materials and methodsFrom October 2020 to May 2021,high resolution CT(HRCT)images with 512×512matrix and ultra-high resolution CT(UHRCT)images with 1024×1024 matrix of patients who had underwent follow-up for pulmonary nodules in our hospital and HRCT images and UHRCT images from phantom were collected retrospectively.A total of 215,000 images(the training set consisted of 9100 HRCT images and 2600 UHRCT images,and the testing set consisted of 5950 HRCT images and 1700 UHRCT images)were included in this study.Based on generative adversarial network,a deep learning framework was constructed for super-resolution reconstruction of CT images.The framework is composed of the main super-resolution branch and the gradient branch,and different loss functions are added to the different branches to optimize the target features of the generated image.The gradient image of HRCT is used as auxiliary information,and the original HRCT image is combined to construct a two-path image generation network to recover richer image details.The peak signal-to-noise ratio(PSNR)and structural similarity index(SSIM)of SRCT images generated by the constructed model,simulated data images,Med SRGAN generated images and GAN-Circle generated images were compared,so as to quantitatively evaluate the quality of SRCT images generated by the network.At the same time,the effectiveness of gradient guidance was verified.ResultsThe PSNR and SSIM of SRCT images generated by the deep learning model in the clinical data set reached 32.84 and 0.872,superior to other methods using simulated data(PSNR:31.46 and SSIM:0.840),superior to Med SRGAN generated images and GAN-Circle generated images(PSNR:31.68,32.31respectively;SSIM:0.841,0.863 respectively).ConclusionThe gradient-guided network model constructed in this study can generate SRCT images via learning HRCT images with 512×512 matrix,and provide a better image quality compared with those models which based on simulated image data,different algorithm model data,or without gradient guidance.Part 2 The Performance Evaluation of Deep Learning Image Super-resolution ModelObjective The images quality of super-resolution CT(SRCT)imaging generated by the deep learning CT image super-resolution model was evaluated subjectively to assess the performance of the model.Materials and methods A total of 29 patients who underwent pulmonary nodules’ follow-up in the Department of Radiology of our hospital from October 2020 to May 2021 were retrospectively enrolled.All patients underwent high-resolution CT(HRCT)with a 512×512 matrix and ultra-high resolution CT(UHRCT)target scanning with a 1024×1024 matrix.SRCT images were generated by using the established deep learning image super-resolution model.Three radiologists with different seniority evaluated the HRCT,UHRCT and SRCT images for noise,streak artifacts,pulmonary nodule edges,clarity of small blood vessels,homogeneity of normal lung parenchyma,and overall image quality.Single-factor repeated measure ANOVA was used to compare noise,streak artifacts,pulmonary nodule edges,clarity of small vessels,homogeneity of normal lung parenchyma and overall image quality between HRCT,UHRCT and SRCT images.Huynh-Feldt and Greenhouse-Geisser,according to different epsilons for sphericity test,had been used as adjustment,and Bonferroni adjustment had been used for post hoc analysis.P < 0.05 had significant statistical difference.Results The scores of noise from high to low were as follow: HRCT,SRCT,and UHRCT(4.49±0.50,3.38±0.74,2.99±0.77,P<0.05).The scores of streak artifact from high to low were as follow: HRCT,SRCT,and UHRCT(4.38±0.51,3.70±0.88,2.87±0.64,P<0.05).The scores of clarity of small vessels from high to low were as follow: SRCT,UHRCT,and HRCT(4.61±0.60,4.03±0.62,2.85±0.56,P<0.05).The scores of overall image quality from high to low were as follow: SRCT,UHRCT,and HRCT(4.63±0.57,4.08±0.61,2.90±0.51,P<0.05).In nodular edge and homogeneity of the normal lung parenchyma,SRCT(4.43±0.60,4.36±0.61 respectively)exhibited no significant difference with UHRCT(4.06±0.91,4.25±0.67 respectively,P>0.05),but exhibited a better performance than HRCT(2.77±0.73,2.97±0.67 respectively,P < 0.05).Conclusion The SRCT constructed by deep-learning CT image super-resolution model in this study had superior performance in noise reduction,artifact removal and image resolution improvement,and a better image quality compared with the HRCT with a 512×512 matrix.Compared with the UHRCT with a 1024×1024 matrix,SRCT images performed better in noise,streak artifacts,clarity of small vessels and overall image quality,and showed similar performance in nodular edge and homogeneity of the normal lung parenchyma.Part 3 Study on Radiomic Features Stability of Deep Learning Super-resolution ImagesObjective To explore the effect of the super-resolution CT(SRCT)image on the stability of radiomics features.Materials and methods A total of 29 patients who underwent pulmonary nodules reexamination in the Department of Radiology of our hospital from October 2020 to May 2021 were ret rospectively enrolled in this study.All 29 patients underwent high-resolution CT(H RCT)with a 512×512 matrix and ultra-high resolution CT(UHRCT)target scannin g with a 1024×1024 matrix.And the corresponding SRCT images were constructed using the established deep learning image super-resolution model.Three radiologists used radiomics software to semi-automatically segment lung nodules from SRCT,H RCT,and UHRCT images in all 29 patients.Then,110 radiomic features were extr acted,including 18 first-order statistics,75 texture features,and 17 shape features.The intragroup correlation coefficient(ICC)was used to calculate the consistency of the indexes on SRCT,HRCT and UHRCT images,and to reveal the influence of SRCT images on the stability of radiomics features.Results The SRCT images generated by the deep learning image super-resolution model had good consistencies with HRCT and UHRCT images in terms of all mentioned radiomics features.Radiomics features with ICC > 0.8 accounts for 70%,and those with ICC > 0.6 accounts for 90%.The mean ICCs of SRCT and HRCT(0.90,0.90,and 0.93 for three radiologists respectively)were higher than that of SRCT and UHRCT,HRCT and UHRCT.Conclusions The SRCT images generated by the deep learning CT image super-resolution model generated in this study were highly consistent with HRCT and UHRCT images in radiomics features,indicating its stability in radiomics features. |
PDF Full Text Request