Deep Learning Based Multiple-Arc VMAT MLC Sequence Prediction in Prostate Cancer
Y Lao1*, N Tong2, W Yang3, K Sheng4, (1) UCLA School of Medicine, Los Angeles, CA, (2) University of Southern California, Los Angeles, CA
Presentations
MO-IePD-TRACK 5-1 (Monday, 7/26/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]
Purpose: To predict sequential multileaf collimator (MLC) apertures and dose for multiple-arc volumetric modulated arc therapy (VMAT), we introduce a deep learning (DL) based model utilizing the beam eye’s views (BEVs) of the given dose distributions in prostate cancer patients.
Methods: Clinical dose distributions and VMAT plans for 70 prostate cancer patients were solicited. All VMAT plans consist of 4 half-arcs, each consisting of 98 control points. A 6-channel 7-level U-NET was used as the main network. The input was the 6 stacked 2D dose distribution slices in the BEV of each gantry angle, spanning from the isocenter to the source with 5mm intervals. The output was the 2 corresponding MLC apertures in each gantry angle, with the associated dose being the intensity. The dataset was divided into 80% for training, 10% for validation, and 10% for testing. 200 epochs were trained using an aggregated loss combining mean squared error and dice coefficient (DC), and the Adam optimizer. For comparison, networks using 4 and 8 input slices per gantry angle were also trained with the same hyperparameters.
Results: On 7 test subjects, the predictions from the 6-channel model (6s-BEV) achieved an averaged MLC apertures DC of 0.73 (± 0.04) and a step-wise dose error (DE) of -4.89% (± 7.98%), normalized to each arc. In contrast, models using fewer (4s) or more (8s) BEV slices as inputs resulted in compromised performance: 0.69 (± 0.04) DC and -10.96% (± 8.35%) DE for 4s-BEV, and 0.71 (± 0.04) DC and -9.06% (± 8.67%) DE for 8s-BEV.
Conclusion: We successfully implemented a U-net for the prediction of multiple-arc prostate VMAT plans. The trained model predicted MLC apertures and the associated dose without involving the time-consuming planning process. The proposed DL model may thus facilitate online applications that benefit from fast planning.
ePosters
Keywords
Taxonomy
IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning
Contact Email