Fully automatic left ventricle segmentation in [Formula: see text]Rb PET/CT Using a semi-supervised nnU-net.
Authors
Affiliations (7)
Affiliations (7)
- Department of Nuclear Medicine and Molecular Imaging, Centre Hospitalier Universitaire Vaudois (CHUV), Lausanne, Switzerland.
- Institute of Informatics, HES-SO Valais-Wallis, Sierre, Switzerland.
- Division of Nuclear Medicine and Molecular Imaging, University of Ottawa, Ottawa, ON, Canada.
- Division of Cardiology, University of Ottawa Heart Institute, Ottawa, ON, Canada.
- Jubilant DraxImage Inc., Kirkland, QC, Canada.
- Department of Cardiology, CHUV, Lausanne, Switzerland.
- Department of Nuclear Medicine and Molecular Imaging, Centre Hospitalier Universitaire Vaudois (CHUV), Lausanne, Switzerland. [email protected].
Abstract
Quantification of myocardial blood flow (MBF) with [Formula: see text]Rb PET/CT requires accurate delineation of the left ventricle (LV). Manual or semi-automated contouring remains time-consuming and error-prone, particularly in hypoperfused myocardium. We developed and validated a fully automatic LV segmentation pipeline using nnU-Net applied to [Formula: see text]Rb PET/CT. A manual, multimodal segmentation protocol integrating dynamic PET and CT was established in a single-center cohort of 40 non-gated PET/CT series (20 patients, rest & stress), including challenging cases with extensive necrosis (median 21%). The resulting ground truth masks were used for five-fold cross-validation, and semi-supervised learning incorporated 805 additional unlabeled dynamic PET series (504 patients). Model performance was compared with an optimized semi-automatic thresholding baseline (35% [Formula: see text]). The nnU-Net significantly outperformed the baseline, achieving a mean Dice of 87.8[85.6, 89.2]% vs 75.1[72.9, 76.9]%, recall 89.1[86.1, 91.4]% vs 82.6[79.1, 85.4]%, and precision 88.1[84.2, 90.4]% vs 70.2[67.2, 73.0]%. The improvement was most pronounced in hypoperfused regions, where recall increased by 20-30% compared to thresholding. Semi-supervised learning modestly enhanced model robustness across both rest and stress acquisitions. A deep-learning-based approach enables fully automatic LV segmentation in [Formula: see text]Rb PET/CT with near-expert accuracy. This framework eliminates manual interaction, supports large-scale MBF quantification, and paves the way for reproducible, high-throughput cardiac PET analysis in clinical and research workflows.