SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Kathleen Gilbert, PhD, BEng
Research Fellow
University of Auckland
Markus Janse, MSc
PhD student
University Medical Center Utrecht
Alistair Young, PhD
Professor
The University of Auckland, New Zealand
Avan Suinesiaputra, PhD
Research Fellow
University of Auckland, New Zealand
Colin Wu, PhD
Mathematical Statistician
National Institutes of Health
David Bluemke, MD, PhD
Professor
University of Wisconsin School of Medicine and Public Health
.jpg "Joao Lima, MD photo")
Joao Lima, MD
Professor of Medicine, Radiology and Epidemiology
Johns Hopkins University
Bharath Ambale-Venkatesh, PhD
Instructor
The Johns Hopkins University
Background:
The Multi-Ethnic Study of Atherosclerosis (MESA) is an observational study with 6814 participants who were aged 45-84 years at the start of the study in July 2000 [1]. The study was designed to follow participants and investigate the development of heart disease. All participants underwent a cardiac magnetic resonance (CMR) examination at baseline, which were contoured using QMASS [2], to derive mass and volume. However, mass and volume measures ignore a large amount of available left ventricular (LV) shape information. We hypothesized that multidimensional shape atlases derived from a machine learning pipeline would show stronger relationships with cardiovascular risk factors rather than standard LV mass and volume indices.
Methods:
A convolutional neural network was trained on 1991 cases, with manual ground truth [2], to identify LV landmarks (mitral valve and right ventricular inserts) using a VGG-16 architecture. The network was then used to automatically generate landmarks on the remainder of the MESA baseline CMR cases. The QMASS contours were extracted and shape models were fitted for both the end-diastole (ED) and end-systole (ES) frames, as shown in figure 1, steps 1-3. After removal of cases with poor quality models, 3897 patient specific models were available for the atlas. The models were then aligned using Procrustes (rotation and translation only) to an average model and principal component analysis applied to the aligned models. Each participant’s scores in each of the principal component analysis modes were extracted from the atlas and used to distinguish differences in shape. Strength of relationships between shape and each risk factor was quantified using the ROC curve obtained from a 5 fold cross validated logistic regression model with the binomial categorical risk factor as the dependent variable and the first 20 modes of each atlas as the independent variables. The area under the curve (AUC) for each model is reported as well as the p-value from DeLong’s test for AUC differences between the atlas model and a reference model using LV mass and volume (ED and ES) only.
Results:
Average errors between manual and VGG-16 landmarks are shown in Table 1. Figure 1e shows the first 3 modes of the ED and ES atlas. The largest mode of variation was size, and the second largest mode was height to width ratio (sphericity). The third mode of the ED atlas was inlet orientation and of the ES atlas was wall thickness. Table 2 shows the results of the logistic regression models. The atlases had significantly stronger relationships (higher AUC) with sex, hypertension medication, high blood pressure and BMI category risk factors than mass and volume only (p<0.001).
Conclusion: Shape atlases derived from a machine learning pipeline may characterize better the effect of cardiovascular risk factors of sex, hypertension and BMI, than standard mass and volume measures.