ESTRO 2022 - Abstract Book

S18

Abstract book

ESTRO 2022

Conclusion We demonstrated a novel hybrid 2D/4D-MRI method for MLC tracking on the Unity MR-linac. Comparable performance to conventional tracking was found, while our workflow provides the ability to reconstruct 4D-MRIs for dose accumulation.

OC-0043 LSTM networks for real-time respiratory motion prediction for a 0.35 T MR-linac

E. Lombardo 1 , Y. Xiong 1 , M. Rabe 1 , L. Nierer 1 , D. Cusumano 2 , L. Placidi 2 , L. Boldrini 2 , S. Corradini 1 , C. Belka 1,3 , M. Riboldi 4 , C. Kurz 1 , G. Landry 5 1 University Hospital, LMU Munich, Radiation Oncology, Munich, Germany; 2 Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Radiation Oncology, Rome, Italy; 3 German Cancer Consortium (DKTK), Munich site, Munich, Germany; 4 Ludwig-Maximilians-Universität München, Medical Physics, Garching, Germany; 5 University Hospital, LMU Munich, Radiation Oncolgy, Munich, Germany Purpose or Objective The MRIdian 0.35 T MR-linac (Viewray Inc, USA) allows for respiratory motion management in clinical practice via a cine MRI gating approach. Although beam gating has been shown to reduce healthy tissue dose compared to ITV or mid- ventilation strategies, it comes with drawbacks such as increased treatment times and the need for patient compliance and is currently not available for all types of MR-linacs. Technologies such as MLC-tracking could address these limitations. However, to perform MLC-tracking the system latency needs to be accounted for. Long short-term memory (LSTM) networks are a type of recurrent neural network which capture temporal dependencies of the input and are therefore ideally suited for motion forecasting. We implemented LSTMs for tumor position prediction based on clinically acquired 4 Hz sagittal 2D- cine MRIs and compared them to a baseline linear regression model. Materials and Methods We collected 556 cine motion videos (106 hours of motion data prior preprocessing) from 88 patients treated at our institution (lung, pancreas, heart, liver, mediastinum). Superior-inferior (SI) motion trajectories of the target centroid were extracted from the contoured videos using in-house software. Breath-holds were excluded, and the curves were normalized and smoothed by performing outlier replacement and applying a moving average filter with a kernel size of three. Patients were divided in training (60%), validation (20%) and testing (20%) sets. The length of the input data windows varied between 2 s and 8 s and was treated as a hyper-parameter. We implemented stateless LSTM networks and optimized them with two different training schemes to predict the future centroid position in 250 ms, 500 ms and 750 ms. For the offline training scheme, we optimized the LSTM on the training set and then applied it to the validation/testing curves. For the offline+online scheme, we loaded the LSTM optimized on the training set but continuously re-trained its parameters on a sliding set of validation/testing data windows in less than 250 ms. We implemented a linear ridge regression (LR) as baseline predictor and determined closed form solution LRs with two different schemes. The offline scheme is analogous to the LSTM while for the online scheme the LR parameters were continuously determined on a sliding set of validation/testing data windows. Results The offline+online LSTM performed best on the testing motion curves (Tab. 1). Both the offline and the online LR models performed worse than the LSTMs, especially for the 500 ms and 750 ms forecasts. Predicted motion data for the best LSTM

Made with FlippingBook Digital Publishing Software