Towards fully automated synthetic ECV quantification: an open-access machine learning-based approach for fast blood draw-free CMR

Extracellular volume (ECV) quantification involves time-consuming multi-step post-processing and a blood draw for hematocrit analysis. This study aimed to develop a fully automated blood draw-free, machine learning-based approach for synthetic ECV assessment for non-invasive assessment of diffuse myocardial fibrosis. We retrospectively evaluated a large clinical cohort of 1092 patients who underwent CMR and ECV measurement at 1.5T or 3T. Participants were divided into training (n = 767) and validation (n = 325) cohorts. Manual contouring of T1 maps was used to iteratively develop a neural network segmentation model, which was then applied for automated analysis. Fully-automated synthetic ECV was calculated using validated sex- and field strength-specific models. Agreement was assessed using Student’s t-test, Pearson correlation, Bland–Altman analysis, and classification analysis. Fully-automated synthetic ECV showed strong correlation with conventional ECV (r = 0.79, p < 0.001), with no significant differences (26.9% ± 4.9% vs. 27.3% ± 6.4%, p = 0.056). Bland–Altman analysis indicated minimal mean difference of 0.4% with moderate limits of agreement (LoA) spanning − 7.24% to + 8.07%, with good agreement for values of up to 35% (mean difference 0.1%, LoA: − 5.38% to + 5.23%). Fully automated synthetic ECV offers a blood-free proof-of-concept for large-scale post-processing, supporting consistent and efficient assessment of myocardial fibrosis in research settings, pending further validation for clinical use at higher ECV ranges.