| Left ventricular hypertrophy(LVH)is a common type of cardiac remodeling and related to adverse clinical outcomes.The most frequent causes of LVH are systemic hypertension,myocardial ischemia,hypertrophic cardiomyopathy(HCM),aortic valve stenosis,chronic renal failure and myocarditis.There are also many other etiologies,such as cardiac amyloidosis,cardiac sarcoidosis,Anderson-Fabry disease,Danon disease,Pompe disease,Noonan’s syndrome,and mitochondrial cardiomyopathy.It is important to identify the primary disease to optimize treatment decision and prognosis.The LVH etiological diagnosis is established mostly upon a comprehensive analysis of clinical information,including history,symptoms,electrocardiogram(ECG),laboratory tests,echocardiography(Echo),and sometimes cardiac resonance imaging(CMR),single photon emission computed tomography(SPECT),positron emission tomography(PET),endomyocardial biopsy and genetic testing,which are complicated,expensive or even invasive.However,due to nonspecific symptoms,nonspecific routine laboratory tests and imaging presentations,and unawareness of the primary disease or family history,it is sometimes difficult to identify the true LVH etiology.Thus,it is clinically significant to seek a simple,cost effective and noninvasive new technique to improve LVH primary disease detection.Transthoracic echocardiography(TTE)is widely used in clinical practice.It is noninvasive,convenient and very effective in discovering cardiac structural and functional abnormalities.Nevertheless,such morphological and functional characteristics from routine TTE scanning is usually nonspecific to identify the LVH etiology.With the booming of computer science in recent years,many efforts are being made to explore computer-assisted diagnosis(CAD)and artificial intelligence(AI)applications in medical research.For instance,radiomics and feature learning in medical image interpretation.With the help of great computing power of computers and mathematical algorithms,it is possible to dig more information out of conventional TTE images,and potentially aid diagnoses.The present study aimed to investigate AI-assisted feature learning based on TTE in differentiating three common LVH entities--hypertensive heart disease(HHD),HCM and uremic cardiomyopathy(UCM).The study consisted of two independent parts.The first part was focused on myocardial texture analysis.Texture refers to the distribution of grayscales in images,and texture features can be quantified through mathematical algorithms.In Echo imaging,pixel gray scales are related to the intensity of reflected soundwaves,and the reflectivity is determined with tissue composition and microstructures.Theoretically,different primary diseases may lead to different myocardial texture patterns in ultrasound imaging.Although such texture variations are subtle and mostly unrecognizable by human visual observation,they might be detected and quantified by an AI program.We enrolled 50 cases for each group in the study.TTE still images of the apical four-chamber(A4ch)were utilized.The interventricular septum(IVS)was outlined manually as the ROI.Then 16 first-order statistics and 60 gray level co-occurrence matrix(GLCM)features were extracted by computer programming,and statistics were performed.Texture features that showed P value < 0.05 by one-way ANOVA were selected to perform inter-and intra-operator correlations and diagnostic accuracy tests.Finally,features with higher area under the receiver operating characteristic curve(AUC)and inter-class correlation coefficient(ICC)> 0.50 were considered more significant and robust,and picked as input variables for AI classification training.Three models of classifiers,namely support vector machine(SVM),logistic regression(LR)and random forest(RF)were tested to find the one with the best diagnostic performance.Main results of the first part study were:(1)Routine Echo parameters were significantly different among the three groups.HCM presented the thickest IVS,,smallest left ventricular(LV)chamber,and highest LV ejection fraction(LVEF).UCM showed the largest LV chamber and lowest LVEF;(2)Texture feature analyses discovered nine parameters,namely Et His,Et Brt,Std,Co V,Skew,Cont7,E11,Hm5 and Et3 that showed significant difference among HHD,HCM and UCM.Diagnostic accuracy tests in two-group comparisons showed that Et Brt,Co V and Std had relatively superior diagnostic performance,with the sum of AUC(AUC-sum)> 4.20.Inter-and intra-operator correlation statistics showed that features with ICC > 0.50 were Et Brt,Std,Co V,Cont7 and Hm5.Thus Etbrt,Std and Co V were finally regarded with both good diagnostic performance and reproducibility,and they were selected for classification validation tests.(3)The best classification result was yielded from SVM model,with the highest AUC-sum of 4.77.The AUCs from SVM were also higher than any single texture feature.The AUCs of Et Brt,Std and Co V individually were 0.61,0.60 and 0.60 in discriminating HHD from(HCM+UCM),while the AUC by SVM was elevated up to 0.70.For the comparison of HHD and HCM.(4)It appeared easy for the computer to distinguish HCM from the others(individual AUC: 0.76 ~ 0.89,AUCSVM: 0.86 ~ 0.91),but relatively difficult to differentiate HHD from UCM(individual AUC: 0.57 ~ 0.77,AUCSVM: 0.68 ~ 0.75).The second part of the study was aimed to perform automated feature extraction and training based on TTE video loops to differentiate HHD,HCM and UCM.Not only gray scale and morphological data,but also myocardial motion data were expected in the AI training and aid diagnosis.Patient recruitment was performed in a retrospective approach by searching the Ultrasound workstation for LVH cases from Jan.1,2018 to Sep.30,2021.Echo images and clinical data were carefully examined to determine LVH etiologies and evaluate image quality.Finally,50 HHD cases,46 HCM cases and 53 UCM cases were enrolled.Then image processing was started from IVS segmentation with a residue U-net(RUNet)model to get the ROI.Then the cardiac motion loops were analyzed frame by frame to extract 46 radiomics features(including intensity,binary and morphological features).Next,by tracing radiomics feature changes with the cardiac motion,we got 24 time domain and frequency domain features.Such that,there were 1104(46 × 24)features for each Echo loop.At last,a modified deep belief network,namely pointwise-gated Boltzmann machine(PGBM)was introduced for feature learning.PGBM was a novel architecture of deep learning,with a switch layer added at the beginning of restricted Boltzmann machine(RBM).The PGMB was linked with a SVM module,fed with the 1104 features with labels,and diagnostic sensitivity,specificity and AUC were calculated.Main results for the TTE video loop study were:(1)Routine Echo parameters again displayed significant difference.HCM group had the smallest LV chamber,thickest IVS,thinnest LVPW and highest LVEF.UCM had the largest LV chamber,but lowest LVEF.Mitral diastolic E/e’ ratio demonstrated declination of LV diastolic function in all three groups,with the most impaired in UCM.(2)Automatic IVS segmentation by RUNet was successful.The tracing results with post-modification reached an accuracy of 98.8%,Dice similarity index of 83.1% and intersection over union(Io U)ratio of 72.8% compared with manual outlining.(3)Deep learning by PGBM resulted in AUCs of 0.70 ~ 0.87 in various comparisons between the groups.It seemed easier to distinguish HCM from the others(AUC: 0.84 ~ 0.87),and somewhat difficult to differentiate HHD and UCM(AUC: 0.70 ~ 0.75).In conclusion,TTE myocardial texture analysis and dynamic feature studying is feasible and potentially valuable for diagnosis.Machine learning enables a comprehensive analysis of multiple texture features and can improve diagnostic accuracy than individual variables.Moreover,deep learning enables automated image segmentation,feature extraction and feature learning.To our knowledge,the study was the first to use time domain and frequency domain features of TTE radiomics.It offered a novel approach to study TTE video loops for AI diagnosis. |
PDF Full Text Request