基于Transformer生成对抗网络模型生成的虚拟CT图像对自发性脑出血后早期血肿的预测研究OA

Objective To investigate the predictive value of virtual CT images generated by a Transformer-based generative adversarial network(TransGAN)model for early hematoma after spontaneous intracerebral hemorrhage(sICH).Methods Patients with sICH who underwent head imaging examinations at the Department of Radiology,Hangzhou First People's Hospital(center 1)and the Department of Radiology,the 903th Hospital of the Joint Logistics Support Force(center 2)from January 2017 to May 2024 were retrospectively and consecutively enrolled.Patients from center 1 were assigned to the training set,and those from center 2 to the test set.Baseline demographic data(age,sex)and imaging data(first head CT examination after onset[baseline CT]and the 24-hour follow-up head CT images)were collected.All head CT images were standardized.The baseline non-contrast CT images and the 24-hour follow-up non-contrast CT images were aligned to the same space through affine registration,serving as paired data for model training.Paired synchronous data augmentation,including random rotation,scaling,and grayscale transformation,was performed on the training set.TransGAN,auto-encoding convolutional neural network(AutoCNN),and conditional generative adversarial networks(cGAN)models were trained by setting hyperparameters and termination conditions.The trained model weights were then loaded into the test set to generate virtual follow-up CT images.The quantitative and subjective evaluation indicators of the virtual follow-up images generated by the three models were compared.With the peak signal-to-noise ratio(PSNR)and structural similarity index measure(SSIM)as quantitative evaluation indictor,PSNR represented the ratio of the maximum possible power of a signal to the power of corrupting noise that affected the fidelity of its representation.A higher PSNR indicated a smaller pixel-level difference between the generated image and the real image,reflecting better image reconstruction quality.SSIM measured the similarity between images based on brightness,contrast,and structural features.A higher SSIM indicated that the generated image was closer to the real image in visual structure.Two neuroimaging physicians with 10 or more years of work experience subjectively evaluated the virtual images generated by the three models using a 5-point Likert scale from three dimensions:hematoma imaging quality,edema imaging quality,and brain parenchyma background quality.A score of 5 indicated that the virtual image was highly similar to the real CT image,with precise hematoma and edema details,and natural and clear brain parenchymal structures;4 indicated definite reference value,with clear hematoma and edema boundaries,and clear brain parenchymal structures;3 indicated basic reference value,with roughly distinguishable hematoma and edema morphology,and visible brain parenchymal structures;2 indicated limited reference value,with blurred hematoma and edema boundaries,and difficult-to-identify brain parenchymal structures;1 indicated almost no reference value,with severe distortion of all structures,making reliable assessment impossible.The weighted Kappa coefficient was used to evaluate inter-rater reliability.Results A total of 311 sICH patients were included,comprising 166 males and 145 females,aged 43-95 years,with a median age of 62(53,72)years.There were 213 cases in the training set and 98 cases in the test set.(1)The age of patients in the training set was higher than that in the test set(P=0.021),while there was no statistically significant difference in gender between the two groups(P=0.851).(2)The PSNR of the virtual images generated by the TransGAN,AutoCNN,and cGAN models in the test set were(26.73±1.11),(22.56±1.53),and(23.54±1.41)dB,respectively.The difference among the three models was statistically significant(F=251.343,P＜0.01).The PSNR of the virtual images generated by the TransGAN model was higher than those of the other two models(both P＜0.01),and the PSNR of the cGAN model was higher than that of the AutoCNN model(P＜0.01).The SSIM of the virtual images generated by the TransGAN,AutoCNN,and cGAN models in the test set were(91.23±1.10)％,(86.78±1.48)％,and(89.32±1.25)％,respectively.The difference among the three models was statistically significant(F=295.232,P＜0.01).The SSIM of the virtual images generated by the TransGAN model was higher than those of the other two models(both P＜0.01),and the SSIM of the cGAN model was higher than that of the AutoCNN model(P＜0.01).(3)The consistency analysis showed that the weighted Kappa coefficients of the Likert scale scores by the two physicians for the virtual images generated by each model were all ≥0.81(0.89 for TransGAN,0.92 for AutoCNN,and 0.82 for cGAN),indicating excellent inter-observer reliability.The Likert scale scores of the virtual images generated by the AutoCNN,TransGAN,and cGAN models in the test set were 3.0(2.0,4.0),4.0(3.0,5.0),and 3.0(2.0,3.0),respectively.The difference among the three groups was statistically significant(x2=251.800,P＜0.01).The Likert scale scores of the virtual images generated by the TransGAN model were higher than those of the other two models(both P＜0.01),and the Likert scale score of the AutoCNN model was higher than that of the cGAN model(P＜0.01).Conclusion The TransGAN model can predict and visualize the early hematoma changes of sICH to a certain extent,providing an imaging reference for a more comprehensive assessment of structural changes in the brain after sICH.