Keywords:
Case-control study, Retrospective, Image registration, Arteriosclerosis, Segmentation, Screening, CAD, CT, Computer applications, Cardiac, Artificial Intelligence, Artificial Intelligence and Machine Learning, Performed at one institution
Authors:
V. Chernina1, M. Pisov1, M. Belyaev1, V. Gombolevskiy1, I. Blokhin1, I. Bekk2, O. Aleshina1, T. Korb1, S. Morozov1; 1Moscow/RU, 2Moscow /RU
DOI:
10.26044/ecr2020/C-10054
Results
We created software for data tagging and export ( Fig. 4 ).
Average tagging time for 1 study was 30 minutes ( Fig. 5 )
A deep learning-based method estimated pericardial contour. First, a shallow network localizes the heart. Next, the region is converted to cylindrical coordinates, and another network determines pericardial contour.
We measured the inter-expert variability via average symmetric surface distance (ASSD). The mean ASSD was 1.88±0.65. The ASSD between the experts and the algorithm was 1.81±0.54 ( Fig. 6 ).
Mean radiologist error was 7,49±5,77 cm3. The mean difference between humans and the algorithm was 10,57±8,61 cm3.
In stress-testing for epicardial fat volume, the correlation coefficient between experts was 98.4%, the mean difference was 8 ml. The human-algorithm correlation was 95,8 % with the mean difference of 12 ml ( Fig. 7 ). Our results are comparable to the results of the algorithm presented by Commandeur F. et al. in 2018 [6]. In addition, the proposed algorithm allows estimating the volume of adipose tissue with high accuracy using non-ECG-gated ULDCT data, which enables its use in screening studies.
Limitations
Due to image noise, other tissues fall within the threshold. We are testing a new denoising algorithm ( Fig. 8).