Author Contributions
Conceptualization, Q.L. (Qiming Liu) and Q.L. (Qifan Lu); methodology, Q.L. (Qiming Liu); software, Q.L. (Qiming Liu); validation, Q.L. (Qiming Liu), Q.L. (Qifan Lu) and Y.C.; formal analysis, Q.L. (Qiming Liu) and Q.L. (Qifan Lu); investigation, Z.T. and Q.W; resources, M.J. and J.P.; data curation, Z.T. and Q.W.; writing—original draft preparation, Q.L. (Qiming Liu) and Q.L. (Qifan Lu); writing—review and editing, Q.L. (Qiming Liu); visualization, Q.L. (Qiming Liu) and Q.L. (Qifan Lu); supervision, J.P.; project administration, M.J. and J.P.; funding acquisition, M.J. All authors have read and agreed to the published version of the manuscript.
Figure 1.
Schematic view of this study. (a) CMR data were collected from HCM, HHD, and HC subjects, and typical mid-ventricle cine images are shown; (b) examples of the volume of interest (VOI) for the MYO, PM, and MYO+PM together in an HC subject (visualization was performed with a 3D slicer); (c) described methods used in the feature selection; (d) shows how the data partition was performed; (e) described ML pipeline: selected MYO and PM features were evaluated for detection and differentiation performance with different ML methods; (f) results were compared between different groups and ML methods, and the evaluation was performed with the ROC curve, calibration curve, and decision curve. HCM, hypertrophic cardiomyopathy; HHD, hypertensive heart disease; VOI, volume of interest; LASSO, least absolute shrinkage and selection operator; ML, machine learning; MYO, myocardium; PM, papillary muscle; LVH, left ventricular hypertrophy.
Figure 1.
Schematic view of this study. (a) CMR data were collected from HCM, HHD, and HC subjects, and typical mid-ventricle cine images are shown; (b) examples of the volume of interest (VOI) for the MYO, PM, and MYO+PM together in an HC subject (visualization was performed with a 3D slicer); (c) described methods used in the feature selection; (d) shows how the data partition was performed; (e) described ML pipeline: selected MYO and PM features were evaluated for detection and differentiation performance with different ML methods; (f) results were compared between different groups and ML methods, and the evaluation was performed with the ROC curve, calibration curve, and decision curve. HCM, hypertrophic cardiomyopathy; HHD, hypertensive heart disease; VOI, volume of interest; LASSO, least absolute shrinkage and selection operator; ML, machine learning; MYO, myocardium; PM, papillary muscle; LVH, left ventricular hypertrophy.
Figure 2.
This figure shows the inclusion and exclusion of LVH and HC subjects. All subjects were obtained from our datasets between 2018 and 2021. For the LVH group, firstly, images with inadequate quality were excluded. Secondly, subjects with myocardial infarction, systemic diseases, or cardiac surgery were further excluded. For the HC group, images with satisfactory quality were matched to the LVH group by sex and age at a 1:2 ratio.
Figure 2.
This figure shows the inclusion and exclusion of LVH and HC subjects. All subjects were obtained from our datasets between 2018 and 2021. For the LVH group, firstly, images with inadequate quality were excluded. Secondly, subjects with myocardial infarction, systemic diseases, or cardiac surgery were further excluded. For the HC group, images with satisfactory quality were matched to the LVH group by sex and age at a 1:2 ratio.
Figure 3.
Performance (AUC, accuracy, precision, and recall) of different models on the detection and differentiation tasks with 3 radiomics subgroups; (a–c) detection task with different ML methods using MYO, PM, and MYO+PM groups, respectively; (d–f) differentiation task with different ML methods using MYO, PM, and MYO+PM groups, respectively. SVM, support vector machine; KNN, K-nearest neighbor; RF, random forest; DT, decision tree; AB, AdaBoost.
Figure 3.
Performance (AUC, accuracy, precision, and recall) of different models on the detection and differentiation tasks with 3 radiomics subgroups; (a–c) detection task with different ML methods using MYO, PM, and MYO+PM groups, respectively; (d–f) differentiation task with different ML methods using MYO, PM, and MYO+PM groups, respectively. SVM, support vector machine; KNN, K-nearest neighbor; RF, random forest; DT, decision tree; AB, AdaBoost.
Figure 4.
Box plot for the best 3 MYO and 3 PM features in differentiation task. The pink and blue boxes represent the range between the first quartile to the third quartile, while the gray dots represent each single patient in the training dataset. Only features in the MYO+PM group for the differentiation task are exhibited.
Figure 4.
Box plot for the best 3 MYO and 3 PM features in differentiation task. The pink and blue boxes represent the range between the first quartile to the third quartile, while the gray dots represent each single patient in the training dataset. Only features in the MYO+PM group for the differentiation task are exhibited.
Figure 5.
ROC curves and calibration curves for differentiation task with MYO, PM, and MYO+PM groups using SVM models. Blue, green, and red lines represent the MYO, PM, and MYO+PM groups, respectively; the AUC for the ROC curves and Hosmer–Lemeshow p-values are denoted in the figure legends. The first row shows the results on the training dataset: (a) ROC curve and (b) calibration curve. The second row shows the results on the testing dataset: (c) ROC curve and (d) calibration curve.
Figure 5.
ROC curves and calibration curves for differentiation task with MYO, PM, and MYO+PM groups using SVM models. Blue, green, and red lines represent the MYO, PM, and MYO+PM groups, respectively; the AUC for the ROC curves and Hosmer–Lemeshow p-values are denoted in the figure legends. The first row shows the results on the training dataset: (a) ROC curve and (b) calibration curve. The second row shows the results on the testing dataset: (c) ROC curve and (d) calibration curve.
Figure 6.
Decision curves for clinical usefulness evaluation. (a) shows the decision curves for the training dataset with the MYO, PM, and MYO+PM groups; (b) shows the decision curves for the training dataset with the radiomics and clinical data groups; (c) shows the decision curves for the testing dataset with the MYO, PM, and MYO+PM groups; (d) shows the decision curves for the testing dataset with the radiomics and clinical data groups.
Figure 6.
Decision curves for clinical usefulness evaluation. (a) shows the decision curves for the training dataset with the MYO, PM, and MYO+PM groups; (b) shows the decision curves for the training dataset with the radiomics and clinical data groups; (c) shows the decision curves for the testing dataset with the MYO, PM, and MYO+PM groups; (d) shows the decision curves for the testing dataset with the radiomics and clinical data groups.
Figure 7.
ROC curve and calibration curve for differentiation task with radiomics and clinical data using SVM model. Red, green, and blue lines represent radiomics, CMR parameters, and radiomics + CMR parameters groups, respectively. The AUC for the ROC curves and Hosmer–Lemeshow p-values are denoted in the figure legends. The first row shows the results on the training dataset: ROC curve (a) and calibration curve (b). The second row shows the results on the testing dataset: ROC curve (c) and calibration curve (d).
Figure 7.
ROC curve and calibration curve for differentiation task with radiomics and clinical data using SVM model. Red, green, and blue lines represent radiomics, CMR parameters, and radiomics + CMR parameters groups, respectively. The AUC for the ROC curves and Hosmer–Lemeshow p-values are denoted in the figure legends. The first row shows the results on the training dataset: ROC curve (a) and calibration curve (b). The second row shows the results on the testing dataset: ROC curve (c) and calibration curve (d).
Table 1.
Demographic and CMR characteristics.
Table 1.
Demographic and CMR characteristics.
Clinical Data Entries | Overall (N = 345) | LVH (N = 230) | HCM (N = 158) | HHD (N = 72) | HC (N = 115) | p-Value 1 |
---|
Demographic and Clinical Features |
Age, year | 51.1 ± 15.7 | 51.4 ± 15.3 | 51.6 ± 15.3 | 50.9 ± 15.4 | 50.6 ± 16.6 | 0.634 |
Male, n (%) | 245 (71) | 165 (72) | 109 (69) | 56 (78) | 80 (70) | 0.171 |
Weight, kg | 71.5 ± 14.3 | 72.4 ± 14.0 | 70.2 ± 12.8 | 77.3 ± 15.4 | 69.8 ± 14.7 | 0.001 |
Height, cm | 168.5 ± 9.0 | 168.3 ± 8.4 | 168.5 ± 8.3 | 170.0 ± 8.5 | 169.1 ± 9.9 | l0.048 |
BMI, kg/m | 25.1 ± 3.9 | 25.4 ± 3.8 | 24.9 ± 3.5 | 26.6 ± 4.2 | 24.3 ± 3.8 | l0.205 |
BSA, m | 1.79 ± 0.22 | 1.80 ± 0.21 | 1.77 ± 0.20 | 1.87 ± 0.23 | 1.77 ± 0.23 | l0.003 |
CMR Parameters |
LVEF, % | 64.1 ± 10.6 | 63.5 ± 11.9 | 67.5 ± 7.4 | 54.8 ± 15.0 | 65.4 ± 7.2 | l < 0.001 |
LVEDV, mL | 135.3 ± 40.7 | 140.2 ± 42.3 | 130.1 ± 29.6 | 162.5 ± 55.7 | 125.5 ± 35.5 | l < 0.001 |
LVEDV index, mL/m | 75.2 ± 19.0 | 77.7 ± 20.3 | 73.8 ± 15.6 | 86.2 ± 26.2 | 70.2 ± 14.8 | l0.002 |
LV mass, g | 123.9 ± 53.2 | 144.3 ± 52.1 | 146.4 ± 54.4 | 139.7 ± 46.6 | 83.2 ± 24.0 | l0.515 |
LV mass index, g/m | 68.8 ± 27.9 | 80.0 ± 27.3 | 82.7 ± 29.1 | 74.1 ± 21.9 | 46.3 ± 9.5 | l0.031 |
Table 2.
The 6 most-important features selected in the MYO and PM groups for HCM vs. HHD according to the Boruta method.
Table 2.
The 6 most-important features selected in the MYO and PM groups for HCM vs. HHD according to the Boruta method.
MYO (N = 6) | Relative Importance | PM (N = 6) | Relative Importance |
---|
gradient GLCM correlation | 37.5 | original shape maximum 2D diameter slice | 10.4 |
original shape sphericity | 4.9 | log-sigma-5-0-mm-3D first-order kurtosis | 3.6 |
original shape elongation | 3.5 | original GLSZM ZoneEntropy | 3.2 |
wavelet-LHL GLCM Imc1 | 3.0 | wavelet-HLL GLCM IMC2 | 3.0 |
log-sigma-5-0-mm-3D glszm ZoneEntropy | 1.3 | log-sigma-2-0-mm-3D GLCM correlation | 2.9 |
wavelet-LHH GLCM MCC | 1.2 | gradient GLCM IDMN | 2.4 |
Table 3.
Performance of selected models on LVH detection and HCM vs. HHD differentiation with MYO, PM, and MYO+PM features.
Table 3.
Performance of selected models on LVH detection and HCM vs. HHD differentiation with MYO, PM, and MYO+PM features.
Group | Feature Number | AUC 1 | Accuracy (%) | Precision | Recall | F1 Score |
---|
LVH detection task (SVM model) |
MYO | 6 | 0.966 2 | 90.4 | 0.903 | 0.829 | 0.853 |
PM | 6 | 0.772 3 | 68.3 | 0.676 | 0.486 | 0.507 |
MYO+PM | 3 + 3 | 0.964 | 91.3 | 0.913 | 0.829 | 0.866 |
HCM vs. HHD differentiation task (SVM model) |
MYO | 6 | 0.875 4 | 82.6 | 0.831 | 0.773 | 0.739 |
PM | 6 | 0.716 5 | 73.9 | 0.811 | 0.182 | 0.308 |
MYO+PM | 3 + 3 | 0.935 | 87.0 | 0.871 | 0.818 | 0.800 |
Table 4.
SVM model accuracy of clinical-based features (including demographic and CMR features) and radiomics-based features on testing dataset.
Table 4.
SVM model accuracy of clinical-based features (including demographic and CMR features) and radiomics-based features on testing dataset.
Evaluation Metrics | CMR Parameters (LVEF + LVEDV Index + LV Mass Index) | Radiomics (MYO + PM) | Radiomics + CMR Parameters |
---|
AUC 1 | 0.774 2 | 0.935 | 0.906 3 |
Accuracy (%) | 71.0 | 87.0 | 85.5 |
Precision | 0.693 | 0.871 | 0.860 |
Recall | 0.409 | 0.818 | 0.818 |
F1-score | 0.474 | 0.800 | 0.783 |
Table 5.
Summary table of main findings in this research.
Table 5.
Summary table of main findings in this research.
Task Name | Matrices | MYO | PM | MYO+PM | CMR | MYO+PM+CMR |
---|
LVH detection | AUC | 0.966 | 0.772 | 0.964 | 0.908 | 0.965 |
Accuracy | 90.4 | 68.3 | 91.3 | 81.7 | 91.3 |
HCM vs. HHD differentiation | AUC | 0.875 | 0.716 | 0.935 | 0.774 | 0.906 |
Accuracy | 82.6 | 73.9 | 87.0 | 71.0 | 85.5 |