External validation of AI-based cardiac MRI quantification in elite endurance and non-endurance athletes

Abstract Background Detecting early/subtle cardiomyopathic features in athletes requires accurate cardiac magnetic resonance (CMR), yet manual segmentation remains variable and non-standardised. Existing AI-based CMR tools, although highly reproducible, are largely trained on non-athlete populations and have rarely been validated in the distinct phenotypes of (elite) athletes. Purpose To validate a commercially available AI-CMR tool in elite athletes against a semi-manual core-lab reference and compare accuracy, bias, and consistency with non-athletes, including sex and endurance versus non-endurance subgroups. Methods Cross-sectional analysis in 160 routinely screened elite athletes (50% women; 50% endurance) and 68 healthy controls who underwent 1.5T CMR. Left, right ventricular (LV/RV) end-diastolic/systolic/stroke volumes (EDV/ESV/SV), LV mass, and maximal LV wall thickness (MWT) were quantified by semi-manual core-lab and an automated AI-CMR algorithm (training set 13% athletes; different cohorts Ref. 1). Quality control of AI output was performed on EDV/ESV snapshots. Primary outcome was difference in absolute-agreement intraclass correlation, ΔICC(2,1), between athletes and controls. Secondary outcomes were volumetric/mass bias (Bland-Altman), consistency ICC(3,1) and performance differences by sex and by endurance versus non-endurance sport types. Results Groups were similar in body size (1.9±0.2 m²; P=0.32). Athletes trained more (20±4 vs 3±4 h/week; P0.001) and had greater LVEDV (120±19 vs 94±15 mL/m²; P0.001). Overall, AI-CMR showed good median absolute agreement across eight CMR metrics (ICC=0.83 IQR: 0.72-0.93) with slightly higher ICCs in athletes (ΔICC +0.04 IQR: +0.02–0.07; pooled Fisher’s z P=0.022; Fig. 1A). In both groups, LVEDV/ESV agreement was excellent (ICC≥0.93) with trivial bias (+4 mL; Fig. 2). RVEDV/ESV agreement was good to moderate (overall ICC=0.88/0.77, respectively) and comparable between athletes and controls (ΔICC=+0.04/–0.04; P0.35) with systematic (+11 mL) and proportional overestimation (slope=+0.11; P0.001). LV mass and MWT had moderate and similar agreement (pooled ICC=0.68 and 0.53, respectively) due to systematic overestimation (+23 g). Despite this, overall consistency was near-excellent (median ICC3,1=0.88, Fig. 1B). For subgroups within athletes, pooled absolute-agreement ICCs did not globally differ by sex (pooled P=0.917) or sport-type (pooled P=0.895). Conclusion This AI-CMR algorithm showed no evidence of performance loss when applied to extreme phenotypes of elite athletes. Agreement with semi-manual quantification was comparable or superior to that in non-athletes. Given high consistency and excellent LV volumetric agreement, AI-CMR can replace manual segmentation for LV volumes and ejection fraction in athletes, enabling scalable, standardised workflows across sexes and sport-types. For absolute RV volumes and MWT, visual review around clinical decision thresholds remains warranted.Figure 1A and B: Forest plotsFor image description, please refer to the figure legend and surrounding text. Figure 2:BA plotsFor image description, please refer to the figure legend and surrounding text.