Automated leaflet assessment and recognition of tricuspid regurgitation with deep neural network

Abstract

Abstract Funding Acknowledgements Type of funding sources: None. Background Timely recognition and intervention of severe tricuspid regurgitation (TR) improve clinical outcomes, especially given advancing transcatheter devices nowadays. We developed deep neural network model to automatically recognize significant tricuspid regurgitation through echocardiography in both assessing color flow and leaflet morphology. Method With careful annotation of corresponded images, we developed three stage models. First model to perform view classification, second joint models to do image processing including 2D image segmentation, leaflet pose estimation, and color image regurgitation flow segmentation, and the third model conduct detection of significant TR. Cross validation was performed stage by stage to confirm performance and model stability. Result The first view classification EfficientNetB5 model reached averaged classification accuracy 94.8%. The second joint models designed in Hourglass structure to achieve averaged 97.1% and 83.3% intersection over union (IOU) for regurgitation flow and chamber segmentation, respectively, and 0.9%-1.6% mean square error for 4 nodes each of three cuspid of tricuspid valve. With the joint of abovementioned post-processing information and original raw image, our model possessed averaged 88% accuracy to detect significant tricuspid regurgitation. Conclusion Our study confirmed that with the utilization of color flow and morphological information, deep neural network model can achieve reliable automatic recognition presence of significant TR.