Rapid Fire Abstracts
Bethlehem Mengesha, MD
Clinical and research fellow
University of Ottawa heart institute, Canada
Rebecca E. Thornhill, PhD
Professor
University of Ottawa, Canada
Suman Prabhakar, MD
Staff Radiologist
University of Ottawa heart institute, Canada
David Ian Paterson, MD, FSCMR
Director of Cardiac Imaging
University of Ottawa Heart Institute, Canada
Gary Small, MD
Medical Director, PET/Nuclear cardiology
University of Ottawa heart institute, Canada
Sharon Chih, MD
Medical director/ Heart Transplant and Mechanical Circulatory Support program
University of Ottawa heart institute, Canada
The prevalence of cardiac amyloidosis (CA) has been increasing, in part due to improved detection from multimodality cardiac imaging. Cardiac magnetic resonance (CMR) tissue characterization is used to diagnosis CA and estimate disease burden. Multiple studies have attempted to differentiate between transthyretin (ATTR) CA and light-chain (AL) subtypes using CMR morphologic features and tissue characteristics [1,2].
Radiomics is an imaging post-processing technique which analyses pixels to derive data on tissue shape and texture and it has been studied in discriminating CA from other hypertrophic phenotypes [3]. The purpose of this study was to train and test a machine learning model developed using radiomic features extracted from routine LGE CMR images to differentiate between ATTR and AL amyloid subtypes.
Methods:
Confirmed cases of CA who underwent a conventional contrast CMR on a 1.5 Tesla scanner (Magnetom Aera, Siemens) were retrospectively identified from a single site. Cardiac function and myocardial tissue characterization (T1, ECV) and LGE quantification were derived using commercial software (CVI42, Circle Inc). LGE images were manually segmented by two experienced readers using 3D slicer and radiomics features were extracted from basal segment short-axis phase-sensitive reconstruction (PSIR) LGE images of left ventricular myocardium using PyRadiomics. The full dataset was randomly stratified into training (n=105) and testing (n=35) sets with respect to amyloidosis subtype. Feature selection was performed on the training set using least absolute shrinkage and selection operator (LASSO) regression. A parsimonious set of radiomic features was then used to train a machine learning algorithm, XGBoost to generate classifiers to discriminate between ATTR and AL subtypes (5-fold CV). Model performance was evaluated on the independent testing set using ROC analysis.
Results:
140 patients (mean age 74±10years, 78% male, 66% ATTR) with confirmed CA were included (Table 1). From a total of 464 radiomic features initially extracted, inter-reader analysis resulted in 210 radiomic features with ICC >0.90. A set of 8 stable radiomic features with nonzero LASSO coefficients were selected for model development. The best XGBoost model showed an AUC ROC of 0.88, (95%CI, 0.75-1) for the identification of AL using morphologic features and radiomic signature compared to model using morphologic features alone (AUC 0.88, 95%CI 0.75-1 and AUC 0.68, 95%CI 0.49-0.87 respectively), DeLong’s test p value=0.03.
Conclusion:
A radiomics analysis of the LGE images on CMR has an incremental value in discriminating between the amyloid subtypes when combined with morphological differences in LV wall thickness, LV size and scar burden.
