CorSeg-CineSAX: An Open-Source Deep Learning Framework for Fully Automatic Segmentation of Short-Axis Cine Cardiac MRI Across Multiple Cardiac Diseases

Le résumé fourni par la source

Abstract Background Segmentation of the left ventricular (LV) myocardium, LV cavity, and right ventricular (RV) cavity on short-axis cine cardiac magnetic resonance (CMR) images is essential for deriving clinical functional parameters, but existing automated tools rarely combine full public availability (source code, trained weights, and a deployment tool) with validation across diverse centers, vendors, and disease categories, particularly at the technically demanding basal right/left ventricular outflow tract (RVOT/LVOT) level. Methods We trained a two-dimensional, slice-by-slice MedNeXt-L model, without region-of-interest cropping, on a private cohort of 2,382 subjects (5 disease categories) from 12 domestic centers, and evaluated it on an independent internal test set (n = 650) and three independent external public datasets (ACDC, M&Ms1, M&Ms2; n = 855 combined), spanning 16 disease categories in total (11 never seen during training). Segmentation accuracy (Dice similarity coefficient [DSC], 95th-percentile Hausdorff distance) was assessed overall, by anatomical slice position (including the basal RVOT/LVOT level), and by disease category, center, vendor, and field strength, alongside agreement between automated and manual clinical functional parameters and benchmarking against human inter-observer variability (three independent observers, 50-subject subsample). Results Mean DSC was 0.908 on the internal test set and 0.883 on the external test set (combined, 0.893; n = 1,505, 31,440 slices). The basal RVOT/LVOT level achieved a mean DSC of 0.904, comparable to the basal (0.910) and mid-ventricular (0.895) levels; the apex was the most challenging position (mean DSC 0.842). Performance on the 11 disease categories absent from training (mean DSC range, 0.874–0.904) was comparable to that on the five categories represented during training (0.881–0.936). Automated LV cavity and right ventricular segmentation matched or exceeded human inter-observer agreement (DSC 0.923 vs. 0.894 and 0.910 vs. 0.906, respectively), while LV myocardium remained somewhat below it (0.847 vs. 0.875). Agreement with manual functional-parameter measurements was strong for LV/RV volumes and LV mass (intraclass correlation coefficient [ICC] ≥ 0.961) and weaker for LV ejection fraction (ICC 0.855) and RV ejection fraction (ICC 0.900), both below the corresponding human inter-observer benchmarks (0.969 and 0.931). Conclusions CorSeg-CineSAX provides an openly released (code, trained weights, and a standalone GUI application), transparently validated framework for fully automatic CMR short-axis segmentation, with performance at the basal outflow-tract level comparable to other anatomical positions and segmentation accuracy for the LV cavity and right ventricle approaching human inter-observer variability.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.