Automated heart segmentation using U-Net in pediatric cardiac CT. (December 2021)

• Record Type: Journal Article
• Authors: Yoshida, Akifumi; Lee, Yongbum; Yoshimura, Norihiko; Kuramoto, Tatsuya; Hasegawa, Akira; Kanazawa, Tsutomu
• Is Part Of: Measurement. Volume 18(2021)
• Publication Date: 2021-12
• Subjects: Segmentation -- Pediatric -- Heart -- Computed tomography -- Deep learning; Detectors -- Periodicals; Measurement -- Periodicals; 530.7
• DOI: 10.1016/j.measen.2021.100127
• Languages: English
• ISSNs: 2665-9174

Abstract

Abstract: This study investigated the usefulness of deep learning methods for segmenting the whole heart region and the cardiac cavity region in pediatric cardiac CT images using U-Net. Dice similarity coefficient (DSC) was used to evaluate the segmentation accuracy by leave-one-subject-out cross-validation. The mean DSC for the whole heart was over 0.95, and analysis of variance among the four age categories (less than one year, 1y to 4y, 5y to 9y, 10y to 14y) showed no significant differences. The mean DSCs for each chamber were 0.78–0.88 when they were trained in a lump. The corresponding DSCs were 0.80–0.85 when they were trained separately. Although the size and shape of the heart varied with age in children, whole heart segmentation using U-Net showed high DSCs in all age categories. Deep learning would become a useful elemental technology in heart segmentation of pediatric cardiology.