| Objective:Treatment for the left main coronary artery disease,three-branch coronary artery disease,or complex coronary artery disease has great challenges.Coronary artery bypass grafting(CABG)is the preferred treatment.Accurately predicting the occurrence of poor prognosis after CABG surgery is of great significance for selecting revascularization strategies and improving prognosis.Currently,the American Society of Thoracic Surgeons(STS)score is mainly used to predict surgical risk.However,this score is primarily used to predict perioperative adverse events.Studies showed that a comprehensive evaluation of left ventricular(LV)function can incrementally predict the long-term risk of death after CABG based on STS score.Therefore,the comprehensive and accurate evaluation of LV function is very important for prognostic prediction of CABG.Echocardiography is the preferred method to evaluate LV function.However,current methods are mostly based on manual measurement and can only be used for gradual evaluation of a single function,which has the limitations of strong empirical dependence,poor repeatability,complex workflow,time consuming and effort.Therefore,there is an urgent need for echocardiographic one-stop intelligent evaluation of LV function,and provides the comprehensive,fast and reliable information for predicting prognostic in patients with CABG.Artificial intelligence technology based on deep learning is developing rapidly in the field of medical imaging,which is expected to provide a new technical means for the one-stop evaluation of LV function.In conclusion,in the first part of this study,we investigated the echocardiographic one-stop intelligent evaluation of left ventricular function based on deep learning.The second part further explored the value of echocardiographic one-stop intelligent evaluation of LV function in predicting the prognosis of CABG.The results of this study will realize the complementary advantages of artificial intelligence and echocardiographic technology,break through the existing limitations of echocardiography,and provide a comprehensive and quick method for evaluating LV function,and thus provide reliable information for prognostic prediction of CABG.Methods:Part I: Based on the two-dimensional apex 3 views,the section classification model and LV global systolic function evaluation model were firstly constructed to fulfill the automatic classification of apex 3 views and automatic measurement of ejection fraction(EF),global longitudinal strain(GLS)and mitral annulus displacement(MAD).Then,the evaluation model of LV regional systolic function was constructed to automatically diagnose the presence of segmental wall motion abnormalities(RWMA).Finally,the evaluation model of LV diastolic function was constructed to automatically diagnose whether the diastolic function was reduced.Ultimately,the one-stop evaluation of LV systolic and diastolic function was realized.Part II: LV systolic and diastolic function was evaluated in the full stack in patients with CABG based on echocardiographic one-stop intelligent method.First,the measurement time and repeatability were compared with the traditional methods by ultrasound expert.Then,the results of ultrasound expert were used as the gold standard to compare the measurement accuracy with that of the resident physician.Finally,the incremental value of echocardiographic one-stop intelligent method in evaluating LV function for predicting prognosis after CABG was investigated.Ultimately,the value of echocardiographic one-stop intelligent method in evaluating LV function before CABG surgery and predicting prognosis after CABG was determined.Results:Part I1.Data setA total of 300 patients were included in the study,and the training and validation of the automatic view classification model,LV global systolic function evaluation model and diastolic function evaluation model were conducted.Among them,150 patients with myocardial infarction and ischemic cardiomyopathy were selected for training and validating the evaluation model of LV regional systolic function.2.View automatic classification and LV global systolic function evaluationView classification model of A2 C,A3C and A4 C has good classification efficiency in the training set and the test set,with the accuracy of 0.97-0.99.LV chambers segmentation model can accurately segment LV chambers and has good segmentation efficiency.The average Dice coefficients of A2 C,A3C and A4 C sections were 0.94,0.93 and 0.95,respectively.The key point detection model can accurately detect the key point sites and has good detection efficiency.The average PCK0.2 of the key point sites on A2 C,A3C and A4 C is0.85,0.72 and 0.89,and the average PCK0.5 is 1.00,0.98 and 0.99,respectively.The model can calculate LV global systolic function parameters in full stack,including LV EF,GLS and MAD.The gold standard was the LV global systolic function results measured manually by ultrasound experts on the machine or commercial software.The predicted value of the model was in good agreement with the gold standard value(ICC 0.80-0.97).3.Automatic evaluation of LV regional systolic functionLV myocardial segmentation model can accurately segment LV myocardium and has good segmentation efficiency.The average Dice coefficients of A2 C,A3C and A4 C segments were 0.87,0.83 and 0.87,respectively.LV myocardial segments model can accurately divide LV myocardium into 18 segments in A2 C,A3C and A4 C.The RWMA identification model showed good diagnostic performance,with an average AUC of 0.83,an accuracy of 0.84,a sensitivity of 0.62,and a specificity of 0.88 for the 5-fold cross-validation.4.Automatic evaluation of LV diastolic functionThe LA/LV chambers segmentation model can accurately segment the LA/LV chambers in A4 C,and has good segmentation efficiency.The average Dice coefficient of LA segmentation is 0.95,and the average Dice of LV segmentation is 0.94.Convex hull algorithm was used to extract LA/LV contour information.LV diastolic function classification model has good diagnostic efficacy,the average AUC of 5 fold cross validation is 0.89,accuracy is 0.91,sensitivity is 0.91,specificity is0.85.Part II1.Data setA total of 125 patients(mean age: 61.2±8.2 years)were included in the study,including 103 males(82.4%).2.Measurement time and repeatability comparisonThe total time spent to evaluate LV function by traditional method was 649±25s.The total time spent to evaluate LV function was 5±0.8s by echocardiographic one-stop intelligent method.Compared with traditional methods by ultrasound experts,the one-stop intelligent method has no operator input and is algorithm-determined,so there is no variability in repeated measurements.3.Accuracy comparisonWith the results of expert evaluation as the gold standard,the LV global systolic function evaluated by the one-stop intelligent method and resident physicians had the good accuracy(ICC about 0.83~0.96;except for the LV end-systolic volume,the percentage error of LOA is less than 30%).The accuracy of the echocardiographic one-stop intelligent method in evaluating the presence or absence of RWMA was better than that of residents(AUCs were 0.86 and0.82,respectively).The accuracy of LV diastolic function evaluated by the echocardiographic one-stop intelligent method was better than that by residents(AUCs were 0.91 and 0.88,respectively).4.Prognostic value15 patients were excluded because the one-stop intelligent method could not be used to evaluate LV function.Among 110 patients with CABG,after a median follow-up of1.95 years,20 patients(18.2%)developed long-term MACE events.STS score and LV global systolic function and diastolic function measured by one-stop intelligent method can predicte long-term MACE after CABG(log-rank test P <0.05).After adding LV global systolic function parameters and diastolic function based on one-stop intelligent method into the original STS model,the C-index of predicting long-term MACE after CABG increased from 0.64 to 0.67~0.70.Adding LV GLS to the original risk classification model based on STS score can significantly improve the classification efficiency of the model(NRI = 0.17,P = 0.03;IDI = 0.005,P = 0.04).Conclusion:1.The deep network model based on two-dimensional apical views can fuifill the full-stack and one-stop intelligent evaluation of LV function,thus providing a new method for the rapid and comprehensive evaluation of LV function.2.The echocardiographic one-stop intelligent method can rapidly,repeatably,and accurately evaluate the LV systolic and diastolic function in a full-stack manner before CABG,and provide the incremental value for predicting long-term prognosis after CABG. |
PDF Full Text Request