SCMR 23rd Annual Scientific Sessions
Quick Fire Session
Patients suspected of hypertrophic cardiomyopathy (HCM) require accurate differentiation from other phenocopy states, such as Anderson-Fabry disease (AFD), cardiac amyloidosis (CA), and clinically unrecognized hypertension (HTN). 3D Myocardial Deformation Analysis (3D-MDA) offers reproducible multi-component phenotypic evaluations that may provide value for machine-learning-based automated classification of these states. We hypothesized that non-linear relationships between regional timing, rate and amplitude of left ventricular deformation exist that are capable of training a machine-learning-based model to classify HCM from other phenocopy states (Non-HCM).
Methods:
One-hundred sixty-three patients with CMR-confirmed LVH were sought from the Cardiovascular Imaging Registry of Calgary (CIROC) meeting objective diagnostic criteria for HCM (N=85) or a non-HCM phenocopy (N=78, inclusive of 30 AFD, 18 CA, and 30 HTN-related cardiomyopathy) adjudicated by comprehensive review of comprehensive LGE-CMR imaging. All subjects had routine, 2D-multiplanar cine SSFP imaging performed using a 3T scanner (Siemens Prisma or Skyra). Routine CMR chamber volumetric measurements were performed in cvi42 (Circle Cardiovascular Imaging). 3D-MDA was blindly performed using locally-developed software (GIUSEPPE) and routinely generated multi-component data provided for machine learning-based modelling. A feed-forward neural network model was adopted, with a 5-neuron hidden layer. Performance was documented as averaged in a 5-fold cross-validation.
Results:
Mean age was 53.1±14.8 years with 68 female subjects (42%) and mean LVEF of 67.8±9.5%. No demographic or routine CMR chamber volumetric measurement achieved an AUC above 70% for classification of HCM versus Non-HCM. Using 3D-MDA data, diagnostic performance for any single global parameter (inclusive of all strain measures) was modest with minimum principal strain performing best with an AUC of 0.64. 3D wall thickness (mean) achieved an AUC of 0.77. Machine learning-based modelling of all available 3D-MDA data (Figure 1) provided an average AUC across the five cross-validation folds of 0.92 (0.83 to 0.96), resulting in an optimal sensitivity and specificity of 0.89 (0.76 to 1.00) and 0.85 (0.80 to 0.88), respectively for the classification of HCM versus Non-HCM (Figure 2).
Conclusion:
Machine-learning-based modelling of regional 3D-MDA data, generated from routine 2D cine CMR, allows for robust differentiation of HCM versus Non-HCM disease etiology. While external validation is required, this methodology shows strong potential for computer-assisted diagnosis using routine, non-contrast cine CMR.
Alessandro Satriano
Senior Cardiac Imaging Software Technician
Stephenson Cardiac Imaging Centre, University of Calgary
Disclosure: Disclosure information not submitted.
Alessandro Satriano
Senior Cardiac Imaging Software Technician
Stephenson Cardiac Imaging Centre, University of Calgary
Disclosure: Disclosure information not submitted.
Yoko Mikami, MD, PhD
Core Laboratory Manager
Stephenson Cardiac Imaging Centre, Libin Cardiovascular Institute of Alberta, University of Calgary
Disclosure: Disclosure information not submitted.
Steven Dykstra, MSc, BEng
PhD Student
Libin Cardiovascular Institute of Alberta, University of Calgary
Disclosure: Disclosure information not submitted.
Jacqueline Flewitt, MSc
Biomedical Engineer
Libin Cardiovascular Institute of Alberta, University of Calgary
Disclosure: Disclosure information not submitted.
Patricia Feuchter, MSc, MRT
Mri research technologist
Stephenson Cardiac Imaging Centre
Disclosure: I do not have any relevant financial / non-financial relationships with any proprietary interests.
Bobby Heydari, MD, MPH
Assistant Professor
University of Calgary
Disclosure: Disclosure information not submitted.
Naeem Merchant, MD
Professor of Radiology
University of Calgary
Disclosure: Disclosure information not submitted.
Andrew Howarth, MD, PhD
Assistant Professor
Libin Cardiovascular Institute of Alberta, University of Calgary
Disclosure: Disclosure information not submitted.
Carmen Lydell, MD
Clinical Assistant Professor, Clinical Co-Director,
Libin Cardiovascular Institute of Alberta, University of Calgary
Disclosure: Disclosure information not submitted.
Gianni Pedrizzetti, PhD
Professor
University of Trieste
Disclosure: Disclosure information not submitted.
Nowell M. Fine, MD, MSc
Clinical Assistant Professor
Libin Cardiovascular Institute of Alberta, University of Calgary
Disclosure: Disclosure information not submitted.
Russell Greiner, PhD
Professor
University of Alberta
Disclosure: Disclosure information not submitted.
James White, MD
Professor of Medicine
Libin Cardiovascular Institute of Alberta, University of Calgary
Disclosure: Cohesic Inc. (Individual(s) Involved: Spouse/Partner, Products/Services: Diagnostic Informatics): Ownership Interest (stocks, stock options, or other ownership interest excluding diversified mutual funds)
