Generative adversarial network for newborn 3D skeleton part segmentation

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 54; no. 5; pp. 4319 - 4333
• Main Authors: Nguyen-Le, Hien-Duyen, Ferrandini, Morgane, Nguyen, Duc-Phong, Tran, Vi-Do, Vo, Hoai-Danh, Nguyen, Tan-Nhu, Dao, Tien-Tuan
• Format: Journal Article
• Language: English
• Published: New York Springer US 01-03-2024; Springer Nature B.V; Springer Verlag
• Subjects: Accuracy; Artificial Intelligence; Artificial neural networks; Bones; Childbirth & labor; Computed tomography; Computer Science; Engineering Sciences; Generative adversarial networks; Image processing; Image reconstruction; Image segmentation; Machines; Manufacturing; Mechanical Engineering; Neural networks; Processes; Three dimensional models; 3D point cloud; Generative adversarial networks; Newborn bones; Skeleton part segmentation; CT scan
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-024-05406-0

Childbirth simulations have been studied in order to predict and prevent difficult delivery issues. The reconstruction of the maternal pelvic model, which consists of a comprehensive fetal model with articulated joints, is important for therapeutic purposes. However, it is difficult and time-consuming to segment the various bones using classical image processing approaches. The aim of this study is to develop and evaluate a generative adversarial network to automatically segment the bony structures of the complete neonatal skeleton. A database of 124 newborn CT images was collected and segmented. Each 3D reconstructed skeleton was divided into 23 distinct bony segments. We proposed the generative adversarial network based on PointNet to perform the automated segmentation directly on the 3D point clouds. Our method was compared to the pointwise convolutional neural network to demonstrate its accuracy and efficiency. The GAN model produced highly accurate results with an IoU of 93.68% ± 7.37%, a Dice of 96.56% ± 4.41% and an accuracy score of 96.72% ± 3.56%, compared to 72.30% ± 5.10% for IoU, 83.82% ± 3.44% for Dice and 84.81% ± 3.25% for accuracy respectively by the pointwise convolutional neural network. In addition, our model behaved better on skeletons in anatomical postures than ones in fetal positions. This study opens new avenues for fast and accurate 3D part segmentation of the newborn 3D skeleton. In the future, further study should focus on segmenting fused bones like vertebrae and integrating the whole articulated skeleton into the maternal pelvic model to simulate complex vaginal delivery and perform associated preventive actions.
Graphical abstract