P123 - ARTIFICIAL INTELLIGENCE FOR AUTOMATED CALCIUM QUANTIFICATION FROM NON-CONTRAST ENHANCED CT IN PATIENTS WITH DEGENERATIVE AORTIC STENOSIS

Friday, October 28, 2022

4:45 PM – 4:55 PM EDT

Location: Room 210

Presenter(s)

Pascal Theriault-Lauzier, MD FRCPC

Interventional Cardiology Fellow
CAIC
Stanford University
Ottawa, Ontario, Canada

First Author(s)

Pascal Theriault-Lauzier, MD FRCPC

Interventional Cardiology Fellow
CAIC
Stanford University
Ottawa, Ontario, Canada

Author(s)

AT

Andy Taylor

McLean, Virginia, United States
JO

Jocelyn Ornelas Munoz

McLean, Virginia, United States
LP

Laura Petruescu

Ottawa, Ontario, Canada
NW

Nadav Willner

Ottawa, Ontario, Canada
RU

Rudy R. Unni, MD, FRCPC (he/him/his)

Cardiology Resident
University of Ottawa Heart Institute
Ottawa, Ontario, Canada
AN

Alex Nantsios

Ottawa, Ontario, Canada
Rebecca Crosier, MD

Ottawa, Ontario, Canada
BT

Bahareh Taji

Ottawa, Ontario, Canada
EW

Emily R Workman

McLean, Virginia, United States
DB

Daniel Berman

McLean, Virginia, United States
DS

David Slater

McLean, Virginia, United States

Senior Author(s)

DM

David Messika-Zeitoun, MD, PhD, FESC

Cardiologist
University of Ottawa Heart Institute
University of Ottawa Heart Institute
Ottawa, Ontario, Canada

Supervisor(s)

DM

David Messika-Zeitoun, MD, PhD, FESC

Cardiologist
University of Ottawa Heart Institute
University of Ottawa Heart Institute
Ottawa, Ontario, Canada

Background: Aortic valve calcification (AVC) is the main pathological process leading to degenerative aortic stenosis (AS). The quantification of AVC is now recommended by guidelines in low-flow low-gradient AS. The last decade has seen the development of convolutional neural networks (CNN), which have found many applications in medicine. The research presented here sought to evaluate if CNN could be used to automatically quantify AVC and localize it for each aortic valve leaflet.

METHODS AND RESULTS: 710 patients who were referred for consideration of aortic valve intervention and who underwent an ECG-gated non-contrast enhanced CT calcium score were included in the AS cohort (mean valve area 0.74±0.18 cm2, mean pressure gradient 44±14 mmHg). The median AVC score was in the severe category (2907 AU), but the range was wide (interquartile range of 2181 AU). An additional 99 patients who underwent a cardiac CT calcium score for another indication were also included in order to provide a control cohort of patients without AVC. Raw images were extracted from the clinical picture archiving and communication system and AVC was manually segmented in all patients with a specialized software. 3D landmarks of the aortic valve complex were annotated including the aortic valve annulus, commissures, and coronary ostia. 80% of images were used for training and 20% were used for validation. A CNN was trained to annotate the aortic valve and reached a 3D positional error between 2.7±1.9 mm and 6.2±3.0 mm for different 3D landmarks. A CNN trained for segmenting AVC voxels from non-AVC voxels was used to calculate the AVC score and reached a mean average error (MAE) of 342 AU. Most of the error was due to calcification of the left ventricular outflow tract (MAE of 212 AU); AVC quantification of the right, left and non-coronary cusps was much more accurate (MAE respectively 46.4, 43.5, and 89.2 AU). During cross-validation, the algorithm classified patient into severe and non-severe AVC (1300 AU for women and 2000 AU for men) with an accuracy of 95.4%, a sensitivity of 99.3%, a specificity of 85.1%, positive predictive value of 94.6% and negative predictive value of 97.9%.

Conclusion: This study demonstrates that a CNN can automatically measure the AVC score with a relatively high accuracy at the level of the aortic root but is less accurate for calcification localized within the left ventricular outflow tract. Despite this, the algorithm can accurately classify patients in severe and non-severe AVC categories.