Abstract
2260
Introduction: Carotid artery disease due to atherosclerosis is a major cause of stroke. Traditionally, we do not examine for pre-symptomatic carotid atherosclerosis, and therapy is predominantly late state repair rather than early phase intervention. The latter requires early detection of atherosclerosis with grading of extent and activity. A novel approach for that is PET/CT with 18F-sodium fluoride (NaF) as tracer, showing micro-calcification long before macro-calcifications become detectable by ultrasound, CT or MRI. However, PET/CT image analysis is done using manual or semi-automated methods, which are time-consuming and require expert image analysis. Recently, utilization of PET studies in the cardiovascular field has been facilitated by introduction of artificial intelligence (AI) models, especially image analysis models. In this project, we developed and tested a convolutional neural network (CNN) approach, a type of AI-based analysis, to segment the common carotid arteries in NaF-PET/CT scans compared with manual segmentation.
Methods: The common carotid arteries were segmented in NaF-PET/CT scans of 29 healthy volunteers and 20 angina pectoris patients. Parameters for segmented volume (Vol) and mean, maximal, and total standardized uptake values (SUVmean, SUVmax, SUVtotal) were analyzed by Bland-Altman Limits of Agreement. Intra- and inter-observer variability of manual segmentation were examined by re-segmentation in 25 randomly selected scans. Repeatability with AI-based assessment of the same scans is 100%. The Sørensen-Dice coefficient (SDC) was calculated to measure the resemblance of CNN and manual segmentations.
Results: An example of the CNN-based and manual segmentation is shown in Figure 1. The Bland-Altman plots are shown in Figure 2. The bias for Vol, SUVmean, SUVmax, and SUVtotal were 1.33±2.06, -0.01±0.05, 0.09±0.48, and 1.18±1.99, respectively, in the left common carotid and 1.89±1.5, -0.07±0.12, 0.05±0.47, and 1.61±1.47, respectively, in right common carotid artery. The manual segmentation took typically 20 minutes per scan compared to one minute with the CNN-based approach. The mean volume deviation at repeat manual segmentation was 14 and 27 percent, respectively, in the left and right common carotids. The mean SDCs for the left and right common carotids are shown in Table 1.
Conclusions: The CNN-based model provided a much faster segmentation of the common carotid arteries. Notably, SUVmean values obtained with the two approaches were very similar suggesting that CNN-based analysis is a promising substitute of slow and cumbersome manual processing.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.