Programa del congreso
Resúmenes y datos de las sesiones para este congreso. Seleccione una fecha o ubicación para mostrar solo las sesiones en ese día o ubicación. Seleccione una sola sesión para obtener una vista detallada (con resúmenes y descargas, si están disponibles).
|
Resumen diario |
| Sesión | |
|
Ju-S5.2-IAMod: Inteligencia artificial y modelado computacional Lugar: Aula 0.01 Presidente de la sesión: Gonzalo Ricardo Ríos Muñoz Presidente de la sesión: Begoña Acha Piñero | |
| Presentación 2 | |
15:15 - 15:30
Automatic segmentation of 3D Left Atrial Meshes in Atrial Fibrillation 1: Universidad Carlos III de Madrid, España; 2: Instituto de Investigación Sanitaria Gregorio Marañón (IiSGM), Hospital General Universitario Gregorio Marañón, Madrid, España; 3: Centro de Investigación Biomédica en Red de Enfermedades Cardiovasculares (CIBERCV), Instituto de Salud Carlos III, Madrid, España Atrial fibrillation is the most common sustained cardiac arrhythmia and often requires catheter ablation guided by electroanatomical 3D mapping. However, the anatomical segmentation of left atrial structures remains manual, which limits reproducibility and clinical integration. We present a deep learning pipeline for automatic segmentation of left atrial surface meshes. Three models were analyzed: MedMeshCNN (edge-based), PointNetGAT (vertex-based with attention), and a two-stage hybrid model combining both. Post-processing and Iterative Closest Point (ICP) registration were evaluated through an ablation study to improve the final performance. The hybrid pipeline achieved the best results in 174 3D left atrial meshes, particularly in complex regions such as the left atrial appendage and pulmonary veins, with a mean accuracy of 0.88, a mean IoU of 0.78, and a mean Dice coefficient of 0.87.
| |
