Artificial intelligence-guided automated measurement of pulmonary vein diameter in patients with atrial fibrillation

Abstract

Abstract Funding Acknowledgements Type of funding sources: Public Institution(s). Main funding source(s): This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education(2022R1I1A1A01071083). This work was also supported by the Korea Medical Device Development Fund grant funded by the Korea government (the Ministry of Science and ICT, the Ministry of Trade, Industry and Energy, the Ministry of Health & Welfare, the Ministry of Food and Drug Safety) (Project Number: RS-2022-00141473). Background In the planning of atrial fibrillation (AF)-related procedures, predicting left atrial (LA) anatomy and pulmonary vein (PV) diameter is important for the effectiveness and safety of the procedures but requires a labor-intensive measurement process. Here, we propose an artificial intelligence (AI) based PV diameter measurement algorithm for the computed tomogram (CT)-based automated PV evaluation. Methods We implemented a mesh-based convolutional neural network for the surface segmentation of four PVs and the LA appendage (LAA) in a 3D LA surface mesh. Our algorithm includes two originative methods of surface depth feature and cohesion loss function to improve the performance. We trained the model with the LA mesh of 210 AF patients’ CT scan and validated the accuracy of surface segmentation and PV diameter with independent 158 samples. Results Using an AI-based automated LA measurement model, we achieved an average Intersection over Union (IoU) of 83.4% and a regional IoU from 78.4 to 87.2 % in 158 LA meshes. When we added the surface depth feature, the IoU was improved by 31.7% compared to the conventional 3D feature. The cohesion loss function reduced the fragmentation rate of the surface label by 3.2%. Post-processed PV diameters did not differ from manually measured left (P=0.56) and right upper PV diameters (P=0.08) but differed in both lower PVs (p<0.001). The eccentricity variance of the PV ostia did not differ between AI-measured and manually measured PVs (P=0.68~0.84). Conclusion We proposed an AI-guided automated algorithm for surface segmentation and PV diameter measurement and validated it at both upper PVs and the eccentricity of the PV ostia. Our algorithm can be applied to the automated sizing of LA appendage and improve labor-intensive manual segmentation.