Room: AAPM ePoster Library
Purpose: Fast adaptive replanning based on real-time images during radiation therapy delivery is difficult with the current treatment planning technology. This work aims to develop a deep learning dose prediction model that can rapidly predict dose based on a segmented image set and be generalized to any tumor site thereby facilitating the first step towards real-time adaptive replanning.
Methods: A deep U-Net was developed to predict a dose distribution based on an image set with segmented organs at risk (OARs) and the planning target volume (PTV). The segmented organs were processed as binary masks and input as individual channels to the network. The model was trained separately using two datasets of head and neck (HN) and gynecological (GYN) cancers, each consisting of 58 training cases and 12 test cases. The loss function was computed as the average mean squared error dose of all OARs and targets. For each test case, the dose qualities of the predicted and manually-generated (ground truth) distributions were compared based on commonly used dose-volume parameters including maximum dose, mean dose, and the volume receiving 50% of the prescription dose (V50) for an OAR and the dose covering 95% volume (D95) for the PTV.
Results: The quality of predicted doses were generally equal or better than the manually generated plans. For HN, the average maximum and mean doses of parotid glands and V50 of spinal cord and parotid glands decreased, while the PTV D95 increased by 1%, as compared to those for the manual plans. For GYN, similar improvement on OAR sparing was observed, while the PTV D95 decreased slightly (by 3%). Once the model was trained, dose prediction time was = 10 s.
Conclusion: The U-Net model rapidly predicts favorable dose distributions based on segmented images of either HN or GYN cancers.
Funding Support, Disclosures, and Conflict of Interest: Funding provided by Manteia Medical Technologies
Not Applicable / None Entered.
TH- RT Interfraction Motion Management: General (most aspects)