Room: AAPM ePoster Library
Purpose: study has shown the feasibility of boosting the accuracy of a simple and fast dose calculation method such as ray-tracing (RT) to that of an accurate one such as collapsed cone convolution/superposition dose through deep learning (DL). In this work, we propose and evaluate three different ways to commission this DL-based dose calculation engine for different linacs.
Methods: we use Acuros dose from the Eclipse treatment planning system as the ground truth dose calculation method. Three commissioning methods were proposed: 1. Use a general RT model for one nominal beam energy and fine-tune the DL model using transfer learning for the target machine. 2. Use pre-trained DL model and commissioning RT model to the beam data of the target machine. 3. Combining the two workflows.
With IRB approval, 112 lung patients treated on Elekta Versa and 29 head and neck patients treated on Varian TrueBeam were retrospectively included in this study. We trained the DL model in Method 1 using 104 lung patients and validated on 8 lung patients (source model). The input of this DL model was RT dose commissioned using one nominal beam energy and CT images and the output was predicted Acuros dose. The model was then transferred to head and neck patients (target model).
Results: 3mm/3% gamma index passing rate of the target model dose is 77.2 %±5.1% while that of the source model is 49.4%±15.3%, which shows that although transfer learning improved model performance, it cannot fully ensure the generalizability. We expect to see better results when we include beam data into the input, as proposed in Method 2 and 3.
Conclusion: proposed three ways to commission a DL-based dose calculation engine and plan to evaluate and will compare them on lung and head and neck cases from Elekta and Varian machines.
Not Applicable / None Entered.