Progetto di ricerca · Università “G. d’Annunzio” / Cyber Surgery
Deep-learning pipeline for automatic localization of thoracic vertebral centroids T1–T11 on DRRs generated from CT volumes, comparing YOLO-Pose and HRNet-W18 with patient-level evaluation splits.
YOLO-Pose MAE≈13.2 px
YOLO→HRNet MAE≈4.6 px
PCK@5≈0.87
Scope
The project addressed automatic thoracic landmark localization as a component for initializing 2D–3D registration in spine-surgery imaging workflows.
Method
DRRs generated from CT volumes with automatically projected 3D-to-2D anatomical landmarks.
YOLO-Pose for global detection and HRNet-W18 for local heatmap-based refinement.
Quantitative evaluation using localization error, PCK and distribution-based metrics.
Results
Mean localization error decreased from approximately 13.2 px with YOLO-Pose to approximately 4.6 px after HRNet refinement. Median error was about 1.9 px, with PCK@5 ≈ 0.87.
HRNet heatmap refinement along the thoracic spine.Error-distribution comparison between YOLO and YOLO→HRNet.
