Purpose: Echocardiographic images play an important role in cardiac diagnosis and functional assessment. Currently, the manual delineation of cardiac substructures can support the qualitative evaluation of cardiac function; however, this time-consuming process prevents real-time assessments during image-guided interventions.. To address this clinical problem, this study proposes a deep learning-based method to rapidly and accurately segment cardiac substructures from echocardiographic images.
Methods: We propose a deep-learning-based echocardiographic image segmentation network, called Echocardiac-Net. In the Echocardiac-Net, a localization subnetwork is used to detect the regions-of-interest (ROIs) of cardiac substructures. Then, a classification subnetwork implements feature enhancement to improve the accuracy of cardiac substructures in a mixed boundary. Finally, a segmentation subnetwork takes the feature map derived from the localization and classification subnetworks to derive the semantic segmentation of each substructure within the respective ROI. A 5-fold cross-validation study was performed using 50 patient echocardiographic datasets. The segmentation accuracy of Echocardiac-Net was evaluated for the endocardium and epicardium of the left ventricle (LV) and left atrium (LA) using the Dice similarity coefficient (DSC) and mean absolute distance (MAD). Manual contours were used as ground truth.
Results: The respective DSC and MAD values for each substructure are as follows: LV endocardium (0.95 ± 0.03 and 0.05 ± 0.06 mm); LV epicardium (0.95 ± 0.03 and 0.05 ± 0.05 mm); and LA (0.89 ± 0.11 and 0.29 ± 1.42). The proposed method can perform the segmentation within 0.3 seconds.
Conclusion: The accuracy of the proposed method was demonstrated by evaluating the differences in the Echocardiac-Net contour centroid and its overlap with ground truth manual contour. The high DSC and sub-millimeter MAD values show that the proposed method has great potential to facilitate myocardial functional assessment in real-time interventional procedures.
Not Applicable / None Entered.