Deep convolutional encoder-decoder network for semantic segmentation of atrial cavity

• M. Saiz Vivó ^[1] ; A. Colomer ^[1] ; V. Naranjo ^[1]
  1. [1] Universidad Politécnica de Valencia

     Universidad Politécnica de Valencia

     Valencia, España

     
• Localización: XXXVIII Congreso Anual de la Sociedad Española de Ingeniería Biomédica. CASEIB 2020: Libro de actas / Roberto Hornero Sánchez (ed. lit.), Jesús Poza Crespo (ed. lit.), Carlos Gómez Peña (ed. lit.), María García Gadañón (ed. lit.), 2020, ISBN 978-84-09-25491-0, págs. 443-446
• Idioma: inglés
• Texto completo no disponible (Saber más ...)
• Resumen

  • Atrial Fibrillation (AF) is the most common cardiac arrythmia and its prevalence increases with the ageing population. It is estimated that there are currently 4.5 million cases in Europe. At present, cardiac ablation is the main treatment procedure for AF. To guide and plan this procedure it is essential for clinicians to obtain 3D anatomical reconstruction models of the atria which are patient specific. The aim of this work is to develop computational models using Deep Learning (DL) techniques for left and right atrium (LA and RA respectively) automatic segmentation from MRI volumetric images. For this, a 3D Dual U-Net algorithm was employed. Multiple models were trained with two different databases; the first database had a large number of training samples (80) and the second database had a small number of training samples (19) and was considered of high variability. The model trained with the second database was capable of accurately segmenting the RA with a Dice coefficient of 0.9160 when fine tuning techniques are implemented. The trained models were also evaluated with a third database for LA segmentation. In the third database evaluation experiment, the network trained with the second database yielded a higher Dice coefficient (0.8515) than the network trained with the first database (0.7715) even though it contained a larger number of training samples. These results suggest that a network trained with a high variability database could improve its generalisation capability and yield good segmentation results when evaluated, without retraining, with an external database.