Inverse treatment planning using deep learning-based organs at risk in radiotherapy for head and neck cancer:...

Radiotherapy (RT) planning for head and neck squamous cell carcinoma (HNSCC) is known to be both challenging and time-consuming. Deep learning (DL)-based auto-segmentation of organs at risk (OARs) may streamline this procedure, although studies have mainly evaluated the geometric accuracy of DL-based contours. Here, we report on a prospective study of inverse treatment planning using DL-based OARs in HNSCC.

This prospective single-center study enrolled 25 patients undergoing definitive or postoperative (chemo-)radiotherapy for HNSCC. Deep learning-based OAR contours were generated using a commercially available auto-segmentation software and rated independently by four radiation oncologists. Clinical RT plans were compared with plans that were re-optimized based on DL-based OARs, while the clinical target volumes (CTV) and planning target volumes (PTV) were kept unchanged. Dosimetric parameters for CTV/PTV and manually delineated OARs were compared using Wilcoxon's signed-rank test.

A total of 259 DL-based contours were rated, with mean scores of > 4 (acceptable with corrections) for all structures. For CTV/PTV coverage, the mean dosimetric differences between clinical and DL-based plans were < 1 Gy for all parameters (D95%, D98%, D2%, Dmean). Similarly, differences between doses to "true" (manual) OARs were largely negligible, with one non-critical outlier observed for the spinal cord. All DL-based treatment plans were deemed clinically acceptable after manual review.

Inverse treatment planning using DL-based OARs in head and neck RT is feasible and results in clinically acceptable plans. Larger studies are required to confirm these results and to move further toward fully automated workflows.