Automated vertebral identification and localization for enhanced radiotherapy patient setup

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Manual identification of vertebral bodies for radiotherapy patient positioning is time-consuming and prone to errors.
• Inaccurate patient alignment can lead to significant treatment mistakes.
• Automated solutions are needed to streamline the radiotherapy setup process.

Purpose of the Study:

• To develop and validate an automated technique for vertebral identification and localization.
• To improve the accuracy and efficiency of image-based patient positioning in radiotherapy.
• To create a widely applicable model for diverse imaging conditions and clinical settings.

Main Methods:

• A retrospective study utilizing an nnU-Net-based model for automated vertebral identification and localization.
• Training involved a large dataset combining public (SpineWeb, Verse19, Verse20, Spine1K) and clinical on-board CT scans.
• A four-step post-processing procedure was implemented to enhance accuracy, considering anatomical variations and abnormalities.

Main Results:

• High identification rates achieved: 97.99% on public datasets and 99.76% on clinical datasets.
• Mean localization errors were low: 1.64 ± 1.23 mm (public) and 1.74 ± 1.36 mm (clinical).
• Section-specific accuracy thresholds were established: 10 mm (cervical), 15 mm (thoracic), and 19 mm (lumbar).

Conclusions:

• The automated approach provides accurate and broadly applicable vertebral identification and localization.
• This technique has the potential to significantly enhance radiotherapy setup workflows.
• The model can serve as a valuable clinical tool for improving patient positioning accuracy.