Joint aortic root segmentation and landmark localization on intraoperative fluoroscopy for TAVI guidance

Accurate intraoperative delineation of the aortic root and localization of key anatomical landmarks during transcatheter aortic valve implantation (TAVI) remain difficult because fluoroscopy provides low soft-tissue contrast and is frequently degraded by motion and overlap from catheters and delivery systems. This study developed and validated a multitask deep learning model for simultaneous aortic root segmentation and landmark localization on fluoroscopic images to support image-guided TAVI. A retrospective dataset of 2,895 fully anonymized fluoroscopic frames from 83 patients who underwent TAVI between 2018 and 2024 was used. Expert annotations included binary masks of the contrast-enhanced aortic root and four anatomical landmarks: two aortic annulus points (AA1, AA2) and two sinotubular junction points near the coronary ostia (STJ1, STJ2). We developed BoundaryAwareMANet (BAMNet), a multitask architecture combining an EfficientNet-V2 encoder, an MA-Net-inspired decoder, a coordinate-aware landmark head, and an auxiliary boundary-guidance pathway. Model performance was evaluated using patient-level five-fold cross-validation. Across five folds, BAMNet achieved Dice $0.916\pm 0.011$ , IoU $0.850\pm 0.018$ , and Surface Dice@4 mm $0.845\pm 0.031$ . Landmark localization reached median and mean errors of $7.64\pm 0.33$ px and $10.02\pm 0.17$ px, corresponding to fold-weighted millimeter errors of 2.03 mm and 2.66 mm after image-specific pixel-spacing conversion. The model produced both segmentation masks and landmark coordinates in a single forward pass, maintaining real-time inference at approximately 63 FPS. Joint segmentation of the aortic root and localization of anatomical landmarks on intraoperative fluoroscopy is feasible.