Diagnostic Study Of Spinal Canal Stenosis Based On Synthetic Lumbar Spine CT Images Generated By U-Net

Objectives: The purpose of this study is to explore the feasibility of the application of neural network based on U-Net to derive synthetic CT images from magnetic resonance images in the diagnosis of lumbar spinal stenosis.Methods: A total of 56 subjects were included in this study,including 39 volunteers without lumbar spinal stenosis and 17 patients with postoperative proven lumbar spinal stenosis.All the subjects underwent T2 WI,T1WI and high-resolution CT scan,and the corresponding synthetic CT images were generated by using the neural network based on U-Net.About 50% of the data in this study is used for training,and the remaining 50% is used to verify the data set.Two senior radiologists will evaluate the consistency of image quality between the synthetic and conventional CT image sets of the same subject.The combination of subjective and objective criteria is used to evaluate the consistency of synthetic and conventional CT image quality.The subjective criteria are mainly the synthetic and conventional CT image quality and the certainty score of focus display.The objective evaluation criteria were mainly the measurement of CT value of the same area of interest delineated by synthetic and conventional CT images,and the calculation of structural similarity ratio of all synthetic CT images.Wilcoxon test was used to evaluate image quality subjectively and paired t test was used to analyze image quality objectively.Two senior doctors independently read and diagnosed the conventional CT and synthetic CT image sets,and took the postoperative results as the reference standard,with the results P < 0.05 as statistically significant.Results: In terms of subjective image quality evaluation,the image quality of conventional CT of 28 subjects was evaluated as "good" by two senior radiologists,but the evaluation of synthetic CT image by two senior radiologists was different.the number of " fair " and "poor" criteria for the quality of synthetic CT images was 1 and 2 for senior radiologists No.1,and 2 and 2 for senior radiologists No.2.The same senior radiologist had no significant difference in the quality evaluation of conventional CT and synthetic CT images,with P values of 0.102 and 0.063 respectively,but showed statistical difference in the quality evaluation of synthetic CT images between the two(P = 0.001).There is a good consensus on the certainty of lesions in synthetic CT images,Cohen kappa is 0.786.In the aspect of objective image quality evaluation,the average value of structure similarity ratio of synthetic CT image is more than 0.9(range: 0-1),and in the aspect of CT value measurement of sketching conventional and synthetic CT image interest area,the p-value is more than 0.05.The diagnostic accuracy of two senior radiologists in the conventional CT image set is 100%,while in the synthetic CT image set,the diagnostic accuracy of two senior radiologists is 82.14%(23 cases)and 78.57%(22 cases),respectively.The p-value of the same senior radiologist in the conventional CT image set and the synthetic CT image set is 0.063 and 0.031 respectively,The p-value of the diagnostic results of two senior radiologists using the synthetic CT image set was 1.Conclusion: 1.The synthetic CT image and the conventional CT image have good consistency in terms of image quality.2.The synthetic CT image of lumbar spine has certain application value in the imaging diagnosis of spinal stenosis.