Background Left ventricular maximum wall thickness (MWT) is central to diagnosis and risk stratification of hypertrophic cardiomyopathy (HCM), but measurement has variation.
Objectives We developed a fully automated machine learning (ML) algorithm for MWT measurement and compared it to international experts using precision (repeatability) on a dataset of HCM patients scanned twice with cardiovascular magnetic resonance (CMR).
Methods Training dataset: Endo- and epicardial end-diastolic contours were derived using a fully-automated convolutional neural network trained on 1,923 independent multi-centre multi-disease cases (14 centres from 3 countries, 10 scanner models, 2 field strengths, with balanced pathologies - health, athletes, myocardial infarction, aortic stenosis, HCM, dilated cardiomyopathy, infiltrative diseases) all segmented by a single expert.
Patients: 60 HCM patients were scanned twice (scan:rescan) in the same session (no biological variability) at different field strengths and vendors (Siemens, GE, Philips) in 3 centres to allow generalizability. The protocol consisted of long axis cines and a short axis (SAX) bSSFP cine stack. Between scans, patients were brought out of the bore, repositioned on the table and re-isocentered.
Wall thickness: MWT was measured in the SAX cine stack in end-diastole (scans A and B) by 11 experts (from 4 continents, 6 countries, 9 centers). For ML performance, the contours were based on a repurposed algorithm used for brain cortical thickness measurement, applying the Laplace equation for all contour points – effectively creating nested smoothly deforming surfaces from endo- to epicardium. We created orthogonal field lines to connect endo-and epicardial points, measured these distances and took the maximum as MWT.
Results 1320 MWT measurements by experts were analyzed. Mean MWT varied significantly from 14.9 mm to 19.0 mm (Δ4.1 mm, p<0.05). MWT measured by ML fell in the middle of the experts (5 read higher, 4 lower, p<0.05). Experts had significantly different test:retest precision, ranging from 1.1±0.9 to 3.7±2.0 mm. ML precision performance surpassed all humans on all measures: precision 0.7±0.6 mm, p<0.05; Bland-Altman limits of agreement (ML 3.7 vs humans average 7.7 mm), and coefficient of variance (ML 4.3% vs experts 5.7–12.1%, p<0.05). Using ML, sudden cardiac death risk prediction would be 1.4 to 3.1 times more precise, and a clinical trial to detect a 2 mm MWT interval change would need 1.3 to 4 (mean 2.3) times fewer patients (beta=0.90, alfa=0.05).
Conclusions ML MWT measurement in HCM is superior to all international experts studied with implications for risk stratification and sample sizes for clinical trials.
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.