Evaluation of automatic pericardial segmentation methods in computed tomography images

Abstract

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