Room: Track 3
Purpose: Pancreas SBRT planning is challenging and time-consuming. This study aims to develop a deep learning framework for the direct generation of IMRT plans via fluence map prediction.
Methods: The framework consists of two sequential convolutional neural networks (CNNs) to predict IMRT fluence and generate deliverable plans: (1) Field dose CNN (FD-CNN) predicts field dose distributions in the PTV and surrounding regions based on patient CT anatomies; (2) Fluence map CNN (FM-CNN) takes the beam’s eye view projection of the predicted field dose as input and predicts the final fluence map. These fluence maps were converted to MLC leaf sequence, and dose was subsequently calculated (model-predicted plans) in Eclipse treatment planning system version 13.6 for plan evaluation. One hundred anonymized pancreas SBRT patients were included in this retrospective study, with training/testing ratio of 85/15. Nine-beam IMRT plans with standardized PTV prescription (33 Gy/5 fractions) and OAR constraints were created by clinical physicists and used as benchmark plans. These paired plans were compared in terms of dosimetric endpoints and monitor units (MUs).
Results: For each patient, the fluence map prediction took 7.1 seconds total on average. PTV mean dose, maximum (0.1 cc) dose, and D95% differences in percentages of prescription were 0.1%, 3.9%, and 2.1%, respectively; OAR mean dose and maximum (0.1 cc) dose differences were 0.2% and 4.4%, respectively. The mean ± standard deviation of total MUs for model-predicted and benchmark fluence maps were 2122 ± 281 and 2265 ± 373, respectively.
Conclusion: A novel deep learning framework for direct plan generation was developed and applied to pancreas SBRT. It produces fluence maps within seconds, thereby bypassing the time-intensive inverse optimization process. The framework could potentially change the paradigm of treatment planning by harnessing the power of deep learning to generate clinically deliverable plans in real time.
Funding Support, Disclosures, and Conflict of Interest: This work is partially supported by an NIH grant (#R01CA201212).