Implementation of deep learning-based auto-segmentation for radiotherapy planning structures: a workflow study at two...

Area of Science:

• Radiotherapy
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning-based auto-segmented contours (DC) for organs at risk (OAR) and clinical target volumes (CTV) have been validated.
• The integration of these validated models into clinical workflows requires performance evaluation and user experience assessment.

Purpose of the Study:

• To evaluate the performance of implemented deep learning-based auto-segmented contour (DC) models in clinical radiotherapy (RT) planning.
• To assess the user experience of Radiation Therapists/Dosimetrists and Radiation Oncologists with these DC models.

Main Methods:

• DC models were implemented at two cancer centers for RT planning in central nervous system (CNS), head and neck (H&N), and prostate cancer.
• Post-contouring surveys assessed required edits and user satisfaction; Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD) compared unedited DCs to approved contours.

Main Results:

• The majority of OAR DCs required minimal subjective and objective edits (mean editing score ≤ 2; mean DSC ≥ 0.90; mean 95% HD ≤ 2.0 mm).
• Mean OAR satisfaction scores were high across CNS (4.1), H&N (4.4), and prostate (4.6) structures.
• Overall CTV satisfaction was also high (4.1), despite limited evaluation.

Conclusions:

• Validated OAR DC models for CNS, H&N, and prostate RT planning demonstrated minimal edits and positive user experience.
• While survey compliance was low, CTV DC models showed high user satisfaction, suggesting their utility as a starting point for patient-specific edits.