Automated contouring of non-contrast echocardiograms result in similar estimates of left ventricular function to manually contoured contrast-enhanced images in chemotherapy patients. (4th February 2022)

• Record Type: Journal Article
• Authors: Akerman, A; Bernard, L; Deschamps, T; Foster, B; Hawkes, W; Mirhadi, E; Piotrowska, H; Sarwar, R; Tetlow, L; Woodward, G; Becher, H
• Is Part Of: European heart journal. Volume 23(2022)Supplement 1
• Publication Date: 2022-02-04
• Subjects: Cardiovascular system -- Imaging -- Periodicals; Heart -- Imaging -- Periodicals; 616.10754
• DOI: 10.1093/ehjci/jeab289.013
• Languages: English
• ISSNs: 2047-2404

Abstract

Abstract: Funding Acknowledgements: Type of funding sources: Private company. Main funding source(s): Ultromics Ltd Introduction: Transthoracic echocardiography (TTE) assessment of left ventricular (LV) function has a central role in early detection and treatment of cancer-therapy related cardiac dysfunction (CTRCD). Contrast TTE is recommended to aid accurate and reproducible LV contouring, but contrast agents are underused and high variability remains. Using machine learning (ML) to automatically contour the LV reduces variability, and may provide comparable clinical benefit to contrast-enhanced TTE. Purpose: Retrospective single site study evaluating agreement between LV volumes and function from manually contoured contrast-enhanced TTE and automated contouring of non-contrast enhanced images. Methods: Adults at risk of developing CTRCD who underwent TTE were recruited. LV volumes (end-diastolic and systolic; EDV and ESV, respectively) and function (ejection fraction; EF) were measured from contrast-enhanced images using manual contouring (MAN-CONT), and non-contrast enhanced images using automation (AUTO-NON). Method comparisons were summarised based on: (i) statistical equivalence (two one-sided t-tests), (ii) systematic difference between methods, via root mean-squared error (RMSE; Deming regression), and (iii) average bias (Bland-Altman). Statistical equivalence bounds were conservatively determined from reproducibility estimates in a similar cohort (EDV, 20 mL; ESV, 15 mL; EF, 5%; REF1), and used to interpret RMSE and bias. Results: The cohort comprised of patients undergoing treatment for breast cancer, lymphoma, and myeloma. Similar body mass indices were seen between groups (median [IQR]: 26.5 [6.7] kg/m2; Kruskall-Wallace test, p = 0.320), but breast cancer patients were younger (55.0 [14.5], 58.5 [23], 60 [9], respectively; p = 0.034). For estimates of LV volumes and function, statistical equivalence, RMSE, and bias are presented in Table1. Average estimates (mean [SD]) of EDV, ESV, and EF were 121 [33] vs. 112 [33] mL, 50 [23] vs. 44 [19] mL, and 60 [9] vs. 61 [8]%, for MAN-CON and AUTO-NON respectively. Comparing between MAN-CON and AUTO-NON in breast cancer patients, all variables were statistically equivalent (Figure1), and while bias was lower than equivalence bounds, RMSE was only lower for EDV and ESV, not EF. For lymphoma patients, estimates of EDV were statistically equivalent, but not EF and ESV. Bias was lower than equivalence bounds for all variables, whereas RMSE was lower for ESV, but not EDV or EF. For myeloma patients, estimates of ESV and EF were statistically equivalent, but not EDV. While bias was lower than equivalence bounds for all variables, RMSE was not lower for any variable. Conclusions: Estimates of LV function from automated contouring of non-contrast TTE are similar to contrast-enhanced TTE (manually contoured), despite poorer image quality. Automated contouring using ML reduces variability, therefore increasing TTE sensitivity, which is critical when clinical management relies on accurate assessment of LV function.