Deep Learning-Based CT-Less Cardiac Segmentation of PET Images: A Robust Methodology for Multi-Tracer Nuclear Cardiovascular Imaging DOI Creative Commons
Yazdan Salimi, Zahra Mansouri, René Nkoulou

et al.

Deleted Journal, Journal Year: 2025, Volume and Issue: unknown

Published: May 6, 2025

Abstract Quantitative cardiovascular PET/CT imaging is useful in the diagnosis of multiple cardiac perfusion and motion pathologies. The common approach for segmentation consists using co-registered CT images, exploiting publicly available deep learning (DL)-based models. However, mismatch between structural images PET uptake limits usefulness these approaches. Besides, performance DL models not consistent over low-dose or ultra-low-dose commonly used clinical imaging. In this work, we developed a DL-based methodology to tackle issue by segmenting directly images. This study included 406 from 146 patients (43 18 F-FDG, 329 13 N-NH 3 , 37 82 Rb images). Using previously trained nnU-Net our group, segmented whole heart three main components, namely left myocardium (LM), ventricle cavity (LV), right (RV) on was resampled resolution edited through combination automated image processing manual correction. corrected masks SUV were fed V2 pipeline be fivefold data split strategy defining two tasks: task #1 #2 components. Fifteen as external validation set. delineated compared with standard reference Dice coefficient, Jaccard distance, mean surface segment volume relative error (%). Task average coefficient internal 0.932 ± 0.033. 15 cases comparable reaching an 0.941 0.018. 0.88 0.063, 0.828 0.091, 0.876 0.062 LM, LV, RV, respectively. There no statistically significant difference among coefficients, neither acquired radiotracers nor different folds ( P -values > 0.05). overall prediction components less than 2%. We acceptable accuracy robust test set nuclear proposed can overcome unreliable segmentations performed

Language: Английский

Deep Learning-Based CT-Less Cardiac Segmentation of PET Images: A Robust Methodology for Multi-Tracer Nuclear Cardiovascular Imaging DOI Creative Commons
Yazdan Salimi, Zahra Mansouri, René Nkoulou

et al.

Deleted Journal, Journal Year: 2025, Volume and Issue: unknown

Published: May 6, 2025

Abstract Quantitative cardiovascular PET/CT imaging is useful in the diagnosis of multiple cardiac perfusion and motion pathologies. The common approach for segmentation consists using co-registered CT images, exploiting publicly available deep learning (DL)-based models. However, mismatch between structural images PET uptake limits usefulness these approaches. Besides, performance DL models not consistent over low-dose or ultra-low-dose commonly used clinical imaging. In this work, we developed a DL-based methodology to tackle issue by segmenting directly images. This study included 406 from 146 patients (43 18 F-FDG, 329 13 N-NH 3 , 37 82 Rb images). Using previously trained nnU-Net our group, segmented whole heart three main components, namely left myocardium (LM), ventricle cavity (LV), right (RV) on was resampled resolution edited through combination automated image processing manual correction. corrected masks SUV were fed V2 pipeline be fivefold data split strategy defining two tasks: task #1 #2 components. Fifteen as external validation set. delineated compared with standard reference Dice coefficient, Jaccard distance, mean surface segment volume relative error (%). Task average coefficient internal 0.932 ± 0.033. 15 cases comparable reaching an 0.941 0.018. 0.88 0.063, 0.828 0.091, 0.876 0.062 LM, LV, RV, respectively. There no statistically significant difference among coefficients, neither acquired radiotracers nor different folds ( P -values > 0.05). overall prediction components less than 2%. We acceptable accuracy robust test set nuclear proposed can overcome unreliable segmentations performed

Language: Английский

Citations

0