Evaluation of automatic pericardial segmentation methods in computed tomography images

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main source(s): Fundação para a Ciência e Tecnologia - FCT. Pericardial fat has been associated with the development many cardiovascular diseases. Non-invasive pericardial assessment may be performed through cardiac computed tomography (CCT), first step being segmentation layers, followed by application threshold based on HU values. Given that manual pericardium is expensive, time-consuming and subject to inter-observer variability, automatic tools have proposed. The purpose this study compare methods experts. Two are compared: commercial solution Siemens deep learning solution. A subset 190 patients were randomly selected from EPICHEART (The influence EPICardial adipose tissue in HEART diseases) (ClinicalTrials.gov: NCT03280433). All data was anonymized prior analysis for purposes study. CCT scans acquired Somatom Sensation 64 slice thickness 3 mm. Manual using 3D Slicer [1] one authors (CS). Segmentation repeated 20 cases CS second observer (FN) assess inter- intra-observer variability. Automatic Cardiac Risk Assessment tool provided within Syngo.via software Healthcare. also developed. Publicly available gathered: CardiacFat dataset [2] which includes two different scanners (Phillips Siemens); OSIC-Konya [3] 87 pulmonary fibrosis patients. used train U-Net architecture [4] 2D (512x512 pixels). network trained batch size 2 rate 0.0001 Dice loss. Excellent intra- interobserver agreement obtained 0.908 0.873 respectively mean absolute distance (MAD) between contours 2.23 3.18mm respectively. however significantly lower performance 0.776 0.773 an MAD 6.47 6.72mm In conclusion, both solutions showed external when compared segmentation, could hinder their applicability clinical studies fat. Particularly methods, generalization datasets known challenging. use publicly vanilla means approach can replicated any researcher, however, training additional more advanced architectures yield improved results enable large studies.