A ResNet34-based multi-level algorithm for visual support of TAVR achieved a classification precision of 92.35%, a recall of 96.02%, and an average localization error of 25.65 pixels.
A novel multi-task neural network architecture using a group structure of landmarks demonstrates high precision and recall for visual support during TAVR procedures.
Abstract The paper presents a new method for visual support of the transcatheter aortic valve replacement (TAVR) procedure, based on a multi-level approach to detecting key points on X-ray images. In contrast to traditional approaches, the authors propose using a group structure of landmarks (anatomical, instrumental, and contour), which allows for taking into account the relationship and joint appearance of points during the operation. For implementation, a multi-task neural network architecture with a single feature extractor and specialized outputs performing visibility classification and regression of point coordinates for each group is proposed. During the experiments, various base models (ResNet18, ResNet34, ResNet50, EfficientNet B7, and VGG16) were compared, among which the ResNet34-based model demonstrated the best performance, achieving a classification accuracy (precision) of 92.35%, a recall (recall) of 96.02%, and an average localization error of 25.65 pixels. To train the model, specialized loss functions (Focal Loss and Wing Loss) adapted to the characteristics of medical data were used. The results confirm the effectiveness and stability of the proposed approach, opening prospects for integrating the developed algorithm into real clinical systems to improve the accuracy and safety of TAVR.
Rusakov et al. (2026) studied Transcatheter aortic valve replacement (TAVR). Multi-level algorithm using a group structure of key points (ResNet34-based model) vs. Other base models (ResNet18, ResNet50, EfficientNet B7, VGG16) was evaluated on Classification accuracy (precision), recall, and average localization error. A ResNet34-based multi-level algorithm for visual support of TAVR achieved a classification precision of 92.35%, a recall of 96.02%, and an average localization error of 25.65 pixels.