Construction And Application Of Coronary Artery Calcification Scores By Non-gated Chest LDCT Based On Deep Learning Model

Part 1 Effect of CT reconstruction kernel on the accuracy of non-gated chest coronary artery calcification score[Objective] Using ECG-gated CT as the gold standard,the effects of different reconstruction kernels on the evaluation of coronary artery calcification by non-gated low-dose CT(LDCT)were discussed.[Methods and Materials] A total of 702 patients who participated in the NELCIN-B3 projects in our hospital were analyzed retrospectively.According to the different scanning schemes,the patients were divided into two groups: group A and group B.Group A was ECG-gated scan and group B was non-gated chest LDCT.According to the different reconstruction methods,group B was divided into three subgroups: B1,B2 and B3,and then reconstructed kernels with smooth,standard and sharp.The image quality was evaluated by the noise value,signal-to-noise ratio(SNR)and contrast noise(CNR)of the four groups of images.At the same time,the radiation doses were recorded,and the coronary artery calcification score(CACS)was performed..The Agatston scores were divided into four standard categories(Agatston scores: 0,1～99,100～399 and > 400).Then the consistency of the CACS results was analyzed by intra-group correlation coefficient(ICC)or Spearman rank analysis was made.At the same time,four groups of images were tested by Kappa to obtain the accuracy of cardiovascular risk stratification.[Results] The noise in group B was significantly higher than that in group A,and the noise in group B1 was the lowest,and the CNR and SNR in group B1 was the highest and that in group B3 was the lowest.There was a good consistency between group B and group A,of which group B1 was the highest(volume ICC=0.932 mass ICC=0.904 Agatston ICC=0.906;P=0.000).The consistency of Agatston score between group B and group A was also high,with Kappa values of 0.757,0.745 and 0.654,respectively,of which the least reclassified patients in group B1.The radiation dose index in group B was significantly lower than that in group A(P < 0.001),and the effective dose was reduced by 41.2%.[Conclusion] Compared with ECG-gated CT,non-gated chest LDCT under different kernels is a reliable imaging method for detecting and quantitatively evaluating CAC.The selection of smooth reconstruction kernel can improve the image quality and accuracy of CACS.Part 2 Construction and preliminary verification of coronary artery calcification scores deep learning model based on non-gated LDCT[Objective] Construct a non-gated coronary artery calcification scores(CACS)deep learning model based on non-gated LDCT data,and preliminarily verify its diagnostic effectiveness under different CT reconstruction kernels.[Materials and methods] The data of 1690 patients screened for three major diseases in our hospital from June 2019 to January 2020 were collected retrospectively,including 338(20%)negative samples and 1352(80%)positive samples,the ECG-gated and non-gated scan datas of all patients were included in pairs.According to the proportion of 7:2:1,the whole data were randomly divided into 1183 training sets(70%,237 negative cases and 946 positive cases)for training cardiac segmentation and calcification extraction,338 verification sets(20%,79 negative and 259 positive)were used to optimize and test the effect during training.finally,the trained model was used to evaluate the performance of the model in 169 test sets(10%,22 negative and 147 positive).At the beginning of the training,ITK-SNAP software was used for image segmentation,and two junior radiologists were double-blindly labeled according to CAC tagging rules,and a senior radiologist reviewed it.Two convolution neural network methods based on improved V-Net structure were used to train and establish a non-gated CACS deep learning model.Then the intra-group correlation coefficient,Pearson linear analysis were used to analyze the consistency of the CACS of the three CT reconstruction kernels on the verification set of 338 cases,and the accuracy of the deep learning model was preliminarily verified.[Results] Under the three CT reconstruction kernels,the sensitivity of the CACS deep learning model was as high as 96.1%,and the specificity was the highest of 93.7%(74/79),and the model had a good consistency compared with the manual CACS(rs=0.905~0.932;P < 0.001),and the difference was not statistically significant.[Conclusion] In this study,we constructed a standardized data set of CAC of non-gated LDCT,and creatively combined two kinds of neural networks based on improved V-Net structure to establish a non-gated CACS deep learning model.This model can obtain stable CACS evaluation results in different CT reconstruction kernels.Part 3 Study on the accuracy of non-gated coronary artery calcification scores deep learning model[Objective] The accuracy of coronary artery calcification score(CACS)of the deep learning model constructed in the previous stage was evaluated in the test sets,and the stability of CACS was evaluated by the deep learning model under different CT reconstruction parameters and different heart rate conditions.[Materials and methods] The pre-trained non-gated CACS deep learning model was used for preliminary clinical evaluation on the test sets of 169 cases.The CT data of all patients were included in pairs.ECG-gated CT was reconstructed by standard kernel,and non-gated chest LDCT was reconstructed by smooth,standard and sharp in turn.Automatic CACS were obtained using a deep learning model,and then all CT data of the patients were measured on the Philips Heart-Beat CS Agatston software.Based on the Agatston score,the cardiovascular risks were divided into 0,199,100,399 and > 400(classified as 0,1,2 and 3 categories).The reliability between model and manual measurement was evaluated by intra-group correlation coefficient(ICC),and the subgroup analysis was performed according to the heart rate of 65 beats / min and the fourth percentile(P75).The accuracy of cardiovascular risk stratification was also obtained by Kappa test.[Results] The CACS and cardiovascular risk stratification results obtained by the deep learning model under different CT reconstruction methods are consistent with the ECG-gated CT(ICC=0.842~0.994;Kappa=0.719~0.838),and can maintain good stability in patients with different heart rates,especially in patients with heart rates higher than 65 beats / min and the fourth percentile(P75).The model can observe a lower cardiovascular risk reclassification rate than manual measurements.[Conclusions] With ECG-gated CT as the gold standard,the non-gated CACS deep learning model constructed in the previous stage can provide accurate CACS,and make heart risk stratification possible in different CT reconstruction kernels.Especially in patients with a heart rate higher than 65 beats per minute and the fourth percentile(P75),the model achieved fewer cardiovascular risk reclassifications.