Comparison of the accuracy and reproducibility of AI based echocardiography and CMR analysis for cardiac function assessment in oncology patients at risk of cardiotoxicity

Abstract Funding Acknowledgements Type of funding sources: None. Background Serial screening for cancer therapy related cardiac dysfunction (CTRCD) demands precise measurement of left ventricular ejection fraction (LVEF) and global longitudinal strain (GLS) to detect cardiotoxicity early and avoid inappropriate treatment breaks. Precision can be maximised by better image quality and more accurate analysis(1). Cardiovascular magnetic resonance (CMR) is gold standard for image acquisition(2,3), but the benefit of fully-automated AI-based analysis is unclear(4). Purpose To compare the precision (test-retest reproducibility) of AI-based analysis with human measurement for quantification of LVEF and GLS using echocardiography (TTE) and CMR. Methods Consented adult oncology patients underwent paired same day repeat scans using TTE and CMR (GE Vivid9 machine and Siemens 1.5T Aera) following EACVI protocols. Manual image analysis was performed by 2 blinded experts (ECHO-PAC and CVI42 software). TTE images were analysed by two FDA-approved fully-automated (unsupervised) AI packages (EchoAI1 and EchoAI2), and CMR images by one FDA-approved AI package (CMRAI1) and one validated fully-automated inline deep-learning algorithm (CMRAI2). Precision was calculated between repeat studies for each modality and analysis technique by the mean absolute difference (MAD), co-efficient of variation (CoV), minimal detectable difference (MDD) and Bland-Altman limits of agreement; compared using Levene’s test. Results 48 patients (median age 52 years, 81% female, 60% breast cancer), underwent same day repeat TTE and CMR (4 scans). Manual median LVEF was lower with TTE than CMR 57% (95%CI 56.2–57.8) vs 61.4% (95%CI 61–67.7), p<0.01. Analysis failed in 8%, 12% and 4% TTE scans using EchoAI1, EchoAI2 and manual techniques. There were no significant differences in the MAD, CoV or MDD in measurements for LVEF (p>0.05) between analysis methods, Table 1. Human analysis was superior to EchoAI1 for measurements of GLS; MAD 1.37 vs 2.71 (p<0.01), MDD 2.5 vs 5.7 (p<0.01) and CoV 6.9% vs 10.6%, (p = 0.03). There were zero analysis failures for analysis of CMR data. The CoV (3.6% vs 4.5%, p = 0.03), MAD (1.8 vs 2.9, p = 0.03) and MDD (4.3 vs 4.8, p>0.5) for LVEF metrics was lower with CMRAI2 compared to human analysis. The CoV (6.2% vs 11.7%, P<0.01), MAD (1.23 vs 1.92, p = 0.03) and MDD (2.4 vs 3.6, p>0.05) for repeat GLS measurements was lower with human analysis compared with CMRAI1. There were no other significant differences, Table 1. Conclusions Measurement precision is higher with CMR than TTE, in keeping with prior studies. Fully automated AI algorithms can provide incremental improvements in LVEF measurement precision for analysing CMR images. AI techniques for analysis of echocardiographic images provide similar precision to expert human readers. With continued development, fully-automated analysis algorithms should improve confidence in CTRCD surveillance imaging for cancer patients, with results generalisable to other clinical indications.