Validation of a Whole Heart Segmentation from Computed Tomography Imaging Using a Deep-Learning Approach

Abstract

Aims: To develop an automated deep-learning-based whole heart segmentation of ECG-gated computed tomography data. Methods: After 21 exclusions, CT acquired before transcatheter aortic valve implantation in 71 patients were reviewed and randomly split in a training (n=55 patients), validation (n=8 patients), and a test set (n=8 patients). A fully automatic deep-learning method combining two convolutional neural networks performed segmentation of 10 cardiovascular structures, which was compared with the manually segmented reference by the Dice index. Correlations and agreement between myocardial volumes and mass were assessed. Results:The algorithm demonstrated high accuracy (Dice score=0.920; interquartile range: 0.906-0.925), and a low computing time (13.4 sec, range 11.9-14.9). Correlations and agreement of volumes and mass were satisfactory for most structures. Six of ten structures were well segmented. Conclusion:Deep-learning-based method allowed automated WHS from ECG-gated CT data with a high accuracy. Challenges remain to improve right-sided structures segmentation and achieve daily clinical application.

Advancements in cardiac structures segmentation: a comprehensive systematic review of deep learning in CT imaging

“…In contrast, only two studies were conducted to segment the PV. Sharobeem achieved a DSC of 0.66 in their 2021 study ( 8 ) and Li achieved DSCs of 0.80 and 0.77 by modifying two models in their 2020 study ( 24 ). Different reasons can explain the low scores.…”

“…Hence, the CT images were expected to have low resolution. Third, inhomogeneous contrast enhancement and beam hardening artefacts were reported to decrease the images’ resolution and affect the segmentation quality ( 8 ).…”

“…The data extracted from each study included information about the study design and purpose (such as the type of cardiac structure being segmented), DL approach (such as method and architecture) and segmentation performance (such as reported DSC values). The classification models, which were employed to assign labels to CT images or specific areas of interest within those images to determine whether particular anatomical structures are present, and the segmentation models, which focus on precisely delineating the boundaries of these structures by labelling individual pixels that allows for the precise partitioning of CT images into distinct regions to represent anatomical structures, were extracted for each study ( 3 , 8 , 10 ). The country of each study was classified based on the location of the first author's primary institution.…”

Deep learning-based workflow for automatic extraction of atria and epicardial adipose tissue on cardiac computed tomography in atrial fibrillation

“…The LA and RA segmentations in these studies achieved Dice coefficients of 0.889 to 0.939 and 0.812 to 0.878, respectively. 13,14,16 We used a dataset of more CT scans with sufficient variability in cardiac morphology; therefore, we achieved superior performance in LA and RA segmentation (LA: 0.960 ± 0.010; RA: 0.945 ± 0.013) compared with these studies. To make the segmentation models practical and meaningful, we also evaluated the volume difference and the correlation between manual and automatic segmentation results.…”

“…Multiple studies have designed different network architectures to extract cardiac structures, including the LA, RA, left ventricle (LV), right ventricle (RV), and LV myocardium from cardiac CT data. [13][14][15][16] However, in these studies on whole-heart segmentation, the datasets used to develop the deep learning models were not aimed at a specific disease, such as AF. Additionally, because AF may lead to structural remodeling, including LA dilatation and tissue fibrosis, 1 reliable automatic segmentation of cardiac structures may not have been achieved in these deep learning models.…”