%Aigaion2 BibTeX export from HES SO Valais Publications
%Monday 31 August 2026 06:14:11 AM

@MISC{,
        author = {Martin Asiain, Maria and Amirian, Mohammadreza and Chevalley, Arthur and Andrearczyk, Vincent and Klein, Ran and DeKemp, Robert and Moulton, Eric and Kamani, Christel H. and Prior, John O. and Jreige, Mario and Depeursinge, Adrien},
      keywords = {82Rb PET/CT, Deep convolutional neural network, Left Ventricle Segmentation},
         month = may,
         title = {Fully Automatic Left Ventricle Segmentation in Dynamic 82Rb PET/CT Using Deep Learning},
          year = {2026},
  howpublished = {Annual Swiss Congress of Radiology (SCR)},
      abstract = {Purpose:
To develop and evaluate a fully automatic deep learning–based method for left-ventricular (LV) segmentation
in dynamic 82Rb PET/CT, building on a robust manual multi-modal LV delineation protocol to
further enable advanced endo-to-epicardial myocardial perfusion analysis.

Materials and Methods:
This retrospective study included 40 non-gated dynamic PET/CT volumes from 20 patients who underwent
82Rb PET/CT examinations for revascularization, each with rest and stress scans. The cohort
included cases with apical thinning and hypoperfused septal and lateral walls. LV contours were manually
delineated across multiple time frames, excluding the blood pool, using CT images to ensure that
contours remained within the myocardium. This dataset was used to train and evaluate a 3D nnU-Net
architecture with 5-fold cross-validation. Performance was assessed using Dice coefficient, precision,
and recall. A simple semi-automatic baseline based on thresholding (35\% of the maximum LV intensity)
was also implemented for comparison.

Conclusion:
Deep convolutional neural networks enable fully automatic LV segmentation in dynamic 82Rb PET and
substantially reduce manual effort. The evaluated nnU-Net model achieves high overall accuracy and
clearly outperforms simple threshold-based methods, including in hypoperfused and thinned regions,
supporting reliable quantification of myocardial perfusion.}
}