Introduction: Noninvasive imaging techniques play a central role in identification, stratification, and follow-up of infectious and inflammatory vascular disease. Manual or semi-automatic quantification of these diseases is time consuming. Besides that, most fully automated analysis software fails to integrate information obtained from hybrid nuclear medicine images like PET/CT. The aim of this study is to develop an accurate, fast, and fully automated algorithm to segment the aorta on low dose CT (LDCT) as a navigator for PET-assessed vascular disease. Methods: Eighteen non-contrast LDCT scans from [ 18 fluor]fluordeoxyglucose ([ 18 F]FDG) PET/CT scans were used. The aorta was manually delineated from aortic valve to iliac bifurcation and used as ground truth. Split, rotation, and noise addition augmentation was used on fifteen scans to create 100 3D patches, which were split into a training (n=80) and validation (n=20) set. Three scans were used as test set. Cropping, normalization, and zero-centering were used for image preprocessing to speed up and improve accuracy of the training. No down sampling was used. A U-Net architecture was modified to process 3D inputs and outputs. The Dice Similarity Metric (DSM) was used to assess algorithm performance. Results: Fully automated segmentation of the aorta was feasible training a modified 3D U-Net on non-contrast LDCT obtained from PET/CT scans. The test set yielded a DSM of 0.846 ± 0.020 (mean ± SD) for the entire aorta. The run time for the test set with the trained 3D U-Net was on average 2 seconds per patient, whereas the manual delineation was in the order of 30 minutes per patient. Conclusions: Comparisons of ground truth and automatic segmentations using the trained 3D U-Net demonstrated excellent concordance. This method can be used for faster and more in-depth analyses of PET tracer uptake in the vascular wall and potentially applied for fast and accurate quantification of inflammatory and infectious vascular diseases.
Read full abstract