Segmentation of the aortic root on angiography images

The article describes an approach based on neural network modeling, which is of great interest for segmenting different anatomical structures during TAVI. Given the complexity of analyzing large volumes of data for cardiac surgeons, the article pays special attention to the automatic analysis of medical data, training, and comparison of modern neural networks. Fifty neural networks were thoroughly examined and tested for predicting aortic root masks (encoders: U-net, U-net++, Linknet, FPN, DeepLabV3+ and decoders: Efficientnet-b0, Efficientnet-b1, Resnext50, Resnet34, Regnetx32). During the training and testing phases, the FPN Efficientnet-b0 cascade architecture demon-strated the best prediction accuracy with metrics IOU 0.771 and Dice 0.870. The conducted study shows that the pro-posed approach based on neural network cascades, which focuses not on detecting key points but on creating segmen-tation masks, allows for real-time prediction of the aortic valve and delivery system location, thereby facilitating correct valve positioning during TAVI.