Document resource
Statement of Contribution: Equal contributions from authors YA, AA, TA, and AB; listed alphabetically.Disclosures: JM is a shareholder in MyCardium AI LtdIntroduction Aortic regurgitation (AR) is usually quantified on phase-contrast (PC) cardiac MRI by manual contouring of the ascending aorta to derive regurgitant fraction (RF). However, this is time-consuming and prone to inter-observer variability. 1CMR® (developed MyCardium AI Ltd) is an artificial intelligence (AI) platform that automatically segments PC images and computes RF. We evaluated its agreement with physician-reported RF, its performance in key subgroups and the potential impact on workflow.Methods We retrospectively screened 8,327 CMR studies at a tertiary centre (Jan 2022–Oct 2024). After excluding studies without an AR RF in the clinical report, missing/poor-quality flow data, or duplicates, 120 cases remained. From these, we chose the first 75 cases of patients with mild (RF <20%; n=20), moderate (20–34%; n=24) and severe (≥34%; n=26) AR by physician report. Anonymised ascending-aorta PC series were processed by 1CMR® to generate RF. We assessed numeric agreement between 1CMR and the physician report using linear regression and Bland–Altman analysis. Pre-specified subgroups were scanner vendor (Siemens vs GE) and aortic valve morphology (tricuspid vs bicuspid). AI inference time was compared with an estimated 240 s/scan for manual analysis.Results Across all 75 cases, AI-derived RF showed excellent correlation (R 2=0.932) with physician RF (figure 1). Bland–Altman analysis demonstrated a small mean bias of –1.95% with 95% limits of agreement –11.8% to +7.86%, indicating good agreement across the AR spectrum (figure 2) and similarity to published human inter-observer variability. Using RF-based categories, AI correctly classified AR severity in 80% of cases; most discrepancies occurred within a few percentage points of thresholds rather than reflecting large numeric errors. AR-severity classification agreement is shown in table 1. Performance remained robust across subgroups (table 2). Median AI inference time was ~6 s/scan versus ~240 s/scan for physician analysis, representing an approximate 40-fold reduction.Conclusion Automated AI analysis of PC-CMR using 1CMR® can rapidly quantify aortic regurgitant fraction with excellent agreement to physician reports and acceptable AR-severity classification, while dramatically reducing analysis time. With appropriate validation and clinician oversight, such tools could help scale quantitative valve assessment and standardise AR follow-up in routine practice.Abstract 244 Table 1AI–human agreement on aortic regurgitation (AR) severityHuman ClassificationAI: MildAI: ModerateAI: SevereMild2000Moderate5172Severe0818Abstract 244 Table 2AI–human RF agreement metrics by subgroupGroupnR2Mean Bias (%)95% limits of Agreement (%)Accuracy (%)Overall750.932-1.95[-11.8, 7.86]80.0Siemens Scanner620.942-2.15[-11.22, 6.93]80.6GE Scanner130.662-1.00[-14.05, 12.05]76.9Tricuspid Aortic Valve630.933-1.62[-11.30, 8.05]81.0Bicuspid Aortic Valve120.670-3.65[-13.92, 6.61]75.0Abstract 244 Figure 1Physician vs. AI RF (n = 75), R2 = 0.932Abstract 244 Figure 2Bland–Altman: physician vs 1CMR RF (n=75), mean bias = -1.95% [–11.8%, 7.86%]