Automatic quantification of left atrium volume for cardiac rhythm analysis leveraging 3D residual UNet for time-varying segmentation of ECG-gated CT

Abstract

Atrial fibrillation (AF) is a heart condition widely recognized as a significant risk factor for stroke. Left atrial (LA) volume variation has been identified as a key predictor of AF, and several researchers have proposed deep learning models capable of quickly providing this measurement by processing computed tomography (CT) or magnetic resonance images. In clinical imaging, time-varying ECG-gated CT offers precise information about LA anatomy and function, which could help in developing personalized treatment plans for AF patients. Furthermore, advancements in time-varying dataset acquisition indicate the potential for expanding the role of CT in the management of AF patients through specialized processing techniques. However, automatic segmentation of the LA across all cardiac phases remains challenging due to significant variations in both anatomical structures and image signals throughout the cardiac cycle. To overcome these challenges, this study presents a comprehensive AI-based framework designed to segment the LA across the entire cardiac cycle and classify patients with AF. Specifically, our framework employs a customized Residual 3D-UNet model to segment the LA from time-varying ECG-gated CT scans and utilizes a One-Class Support Vector Machine (OCSVM) to distinguish patients in sinus rhythm (SR) from those with AF. A dataset of 93 time-varying ECG-gated CT scans was retrospectively collected: 60 patients were used for the segmentation task, while 33 patients were used for the classification task. The Residual 3D-UNet model demonstrated high accuracy, achieving a mean Dice score of 0.94, with consistent precision (94.45%) and recall (94.83%) across ten cardiac phases. The OCSVM achieved 78.7% accuracy with high specificity (86.3%), effectively minimizing the risk of misclassifying AF as SR, although sensitivity was lower at 70%, demonstrating the potential of automated segmentation and rhythm classification, providing a potential valuable tool for AF diagnosis.

Keywords: Atrial fibrillation; Left atrium; Segmentation; Time varying ECG-gated CT; Volume variation analysis.

Graphical abstract

Fig. 1

Diagram illustrating the workflow followed for this study.

Fig. 2

Ground truth manually segmented in the axial (x-y), coronal (x-z), and sagittal (y-z) planes of a CT scan and the corresponding 3D visualization. The ROI used for the SNR calculation is also shown in the coronal view.

Fig. 3

(a) Model architecture; (b) convolutional residual sub-unit and (c) up-convolutional residual sub-unit. e1, e2, and e3 refer to the encoding stages, while d1, d2, and d3 indicate the decoding stages.

Fig. 4

(a) 3D model and (b) LA antero-posterior diameter obtained from the segmentation of the LA by the trained Residual 3D-UNet.

Fig. 5

Training loss (T_loss) and validation score (V_score) trends over 500 epochs. The bands represent the relative Interquartile Range (IQR) of the two metrics.

Fig. 6

(a) An example of the distances between manual and automatic segmentation for all phases of the cardiac cycle, shown in a double view; (b) magnification of the worst case (phase 0%). The negative and positive values represent the node-to-node errors in terms of underestimation and overestimation of the AI model with respect to the manual one, respectively. The histogram (c) quantifies the number of nodes with the associated node-to-node error for phase 0%.

Fig. 7

Representation of the cubic splines obtained from ten volumes across the cardiac cycle of a SR (violet) and AF (orange) patient.

Fig. 8

a-c) Histograms of the distribution of LAEI, LAEF, and AP features. d-f) Box plots showing the distribution of LAEI, LAEF, and AP measurements for AF and SR patients. The LAEI and LAEF are expressed as percentages, while the AP is measured in millimeters (mm). Each box plot displays the median, interquartile range, and any outliers for each parameter.

Fig. 9

Distribution of decision function scores for the SR and AF patients computed by the OCSVM algorithm. The black dashed line indicates the primary decision boundary, corresponding to the origin of the hyperplane.