Main Session
Sep
29
PQA 07 - Head and Neck Cancer, Lung Cancer/Thoracic Malignancies, and Nursing and Supportive Care
Presenter(s)
William Green, MD - University of Pennsylvania Perelman School of Medicine, Radnor, Pa
W. R. Green1,2, R. McBeth1, E. Berlin1, M. Costea3, G. Temiz4, B. Ungun4, R. Vauclin4, E. Mengin4, N. Paragios4, A. Chakrabarti4, and P. Maury5; 1Department of Radiation Oncology, University of Pennsylvania, Philadelphia, PA, 2University of Pennslyvania, Radnor, PA, 3TheraPanacea, Lyon, France, 4TheraPanacea, Paris, France, 5Gustave Roussy, Villejuif, France
Purpose/Objective(s):
Automation in radiotherapy planning aims to improve efficiency, plan quality, and reduce interobserver variability. Two-step auto-planning (AP) workflows combining dose prediction and dose mimicking have shown promising results. This study developed and evaluated two end-to-end automated VMAT planning methodologies for head and neck radiotherapy in oropharyngeal cancer.Materials/Methods:
Two AP methodologies were implemented. AP-Method1 used a dose prediction model based a Convolutional Neural Network (CNN) trained on 208 datasets from two centers. AP-Method2 used a published analytical dose prediction formula [1] to generate a patient specific dose prediction. Both methods used an identical dose mimicking component. External validation was performed on 12 patients. For each case, two automated VMAT treatment plans were generated for a VersaHD linear accelerator, prescribing 70Gy, 60Gy and 54Gy to high-, intermediate-, and low-risk targets in 33 fractions. The automatic plans were calculated in RayStation treatment planning system. Dosimetric results were compared with manual clinical plans using guideline-based criteria and analyzed with the Wilcoxon rank-sum test. Qualitative evaluation was performed by one medical physicist and one radiation oncologist assessing clinical acceptability and plan preference.Results:
The mean generation time was approximately 10 minutes per plan. Both AP methods produced clinically comparable or superior results versus manual plans, with similar PTV coverage and improved parotid sparing. Significant differences for AP-Method1 were observed for PTV60Gy, right parotid, and optic nerves. For AP-Method2, differences were seen for mandible, right parotid, and optic nerves. Clinical acceptability was 79% for AP-Method1 and 92% for AP-Method2. In the event of two acceptable cases, reviewers were asked for their preferred plan. Plan preference was 58% for AP-Method1 and 38% for AP-Method2, with full agreement in 7/12 cases. Reviewer assessments indicated that increased parotid sparing in automated plans occasionally corresponded with minor reductions in target coverage. This observation underscores the balance between OAR and target coverage in head and neck planning and reflects differing clinical strategies among experts.Conclusion:
Both automated VMAT methodologies generated high-quality, clinically acceptable plans with minimal human intervention. The strong dosimetric performance, high acceptability rates, and short generation time demonstrate feasibility and promising clinical performance of end-to-end automated planning for oropharyngeal cancer. Implementation of such workflows has the potential to standardize plan quality, reduce variability, and significantly enhance efficiency in routine clinical practice. References: [1] Munshi A, et al. Dose fall-off patterns with VMAT and 3D-CRT including the “organ at risk” effect. Cancer Radiother. 2019;23:138–146.