%0 Journal Article %A Sofie Skovrup %A Reza Piri %A Lars Edenbrandt %A Mans Larsson %A Olof Enqvist %A Oke Gerke %A Poul Flemming Hoilund-Carlsen %T Global cardiac atherosclerotic burden assessed by fast automated artificial intelligence-based heart segmentation in 18F-sodium fluoride PET/CT scans: head-to-head comparison with manual segmentation %D 2021 %J Journal of Nuclear Medicine %P 29-29 %V 62 %N supplement 1 %X 29Introduction: Artificial Intelligence (AI)-based models are increasingly being used to improve and speed up clinical decision-making and research processes. This study compared a fast AI-based method for segmenting the heart in positron emission tomography/computed tomography (PET/CT) scans with manual segmentation to assess global coronary atherosclerosis burden in its early stages. Methods: Heart segmentation was based on a convolutional neural network (CNN) used in 18F-sodium fluoride PET/CT scans of 29 healthy control subjects and 20 angina pectoris patients and compared with data obtained by manual segmentation in the same scans. The CNN was trained on a separate dataset and trained to include the atria and ventricles and exclude the great vessels. The corresponding manual approach to define the cranial boundary of the heart was the outer border of the right pulmonary artery in the mid-sagittal plane. Obtained parameters for segmented volume (Vol) and mean, maximal, and total standardized uptake values (SUVmean, SUVmax, SUVtotal) were analysed by Bland-Altman limits of agreement. Reproducibility with AI-based assessment of the same scans is 100%. Reproducibility of the manual approach was examined by manual re-segmentation in 25 randomly selected scans. Results: An example of the AI-based and manual segmentation is shown in Figure 1. Bland-Altman plots are shown in Figure 2. Mean values (± standard deviation) obtained with AI-based and manual segmentation were: Vol 579.92±162.08 vs. 731.50±158.38 (p<0.001), SUVmean 0.68±0.15 vs. 0.68±0.15 (p=0.55), SUVmax 2.59±0.87 vs. 2.87±1.11 (p=0.03), SUVtotal 394.05±127.41 vs. 500.93±150.68 (p<0.001). Corresponding values for bias were 151.58±72.91, 0±0.02, 0.28±0.87 and 106.88±62.91, respectively. The manual segmentation lasted typically 30 minutes per scan vs. about one minute with the CNN-based approach. The maximal volume deviation at repeat manual segmentation was 8 percent. Conclusion: The CNN-based method was a faster approach and provided values for Vol and SUVtotal that were about 20 percent lower than the manually obtained values, whereas SUVmean and SUVmax values were comparable. This AI-based segmentation approach may offer a more reliable and much faster substitute for slow and cumbersome manual segmentation. The differences between the AI-based and manual segmentations were mainly due to different approaches to define the cranial border of the heart. Supporting data: Figure 1. Axial (a), coronal (b) and sagittal (c) reconstruction of manual (upper panel) and CNN-based (lower panel) heart segmentation in the same patient. Figure 2. Bland-Altman plots for differences between Vol (A), SUVmean (B), SUVmax (C), and SUVtotal (D) obtained by manual and CNN-based segmentation (Manual - CNN) plotted against average ((Manual + CNN )/ 2) in the heart (n=49). The estimated bias of one method relative to the other is the mean difference between values obtained by the two methods shown as a thick black horizontal line in the center with its 95% confidence interval (green shade), whereas the limits of agreement are indicated by the thin black horizontal line lines with their 95% confidence interval (blue shades). %U