Automatic part segmentation for full newborn skeleton-articulated geometries using geometric deep learning and 3D point cloud
Résumé
The development of the maternal pelvis model including a detailed foetal model with articulated joints is of great clinical relevance. The objective of the present study is to propose an automatic and fast segmentation workflow of the full newborn skeleton using geometric deep learning. Computed tomography scans of 124 newborn were retrieved and manually segmented. PointNet++, a geometric deep learning algorithm, was trained to perform automatic segmentation on the 3D point clouds of the reconstructed skeletons. This method was compared with the k-means clustering approach. The PointNet++ model provided highly accurate results, with an accuracy of 95.7 ± 4.7% and an IoU of 91.7 ± 7.9%, while k-means clustering provided unsatisfactory results (Accuracy = 74.6 ± 3.7% and IoU = 59.8 ± 4.7%). This study provided a powerful and accurate automatic segmentation workflow for the full newborn skeleton.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |