Poster Poster Program Therapy Physics

Spatiotemporal Deep Learning of VMAT Direct Aperture Optimization Dynamics

Abstract
Purpose

Dose prediction models operate in the CT dose domain and require a conventional optimization step to translate the predicted dose into deliverable parameters in beam’s-eye view (BEV). To address the taxing direct aperture optimization (DAO) process, directly leverage BEV information, and reduce planning time, we propose a spatiotemporal deep learning framework that represents all LINAC control points as a frame, and models the progression of optimization iterations as frame-to-frame evolution in a video sequence. Under this formulation, DAO is cast as a frame prediction problem.

Methods

Thirty prostate plans were generated in matRad, with the full workflow performed. During DAO, aperture shapes and weights for all control points were recorded per iteration. Spatial information of all control points for each iteration was then preprocessed into 2D frames. A convLSTM was then built, capturing spatial correlations between control points and modeling dependencies across iterations, enabling the network to learn dynamics. The model was trained using a mean squared error loss between the predicted and actual aperture configuration.

Results

The model demonstrated strong performance in predicting MLC leaf positions, achieving an R² of 0.8603, RMSE of 3.248 mm, and MAE of 1.755 mm for the left leaf positions, with similar results for the right leaf positions. For beam weight prediction, the model achieved an R² of 0.5950, RMSE of 1.375 mm and MAE of 0.383 mm, indicating moderate correlation with relatively low absolute prediction error. When the aperture shapes and weights were used to initialize the DAO in MatRad, a mean reduction in iterations of 13.7% and a mean reduction in optimization time of 5.5% were observed.

Conclusion

This work presents promising results in re-imagining DAO as a sequential spatiotemporal deep learning problem. Future work will validate alternative up-to-date sequential models, such as Video Vision Transformer.

People

Related

Similar sessions

Poster Poster Program
Jul 19 · 07:00
Python-Based Automation Framework for Annual Machine QA Data Archiving In Qatrack+

Annual water-tank measurements help ensure beam characteristics remain consistent with commissioning baselines. However, the lack of a standardized processing workflow and decentralized data storage makes it difficult to analyze...

Syed Bilal Ahmad, PhD
Therapy Physics 0 people interested
Poster Poster Program
Jul 19 · 07:00
User Expectations and Current Availability of HDR Brachytherapy Audits In Europe

The aim of this work was to evaluate the need to implement more dosimetric audits in high‐dose‐rate brachytherapy (HDR-BT) in Europe and to identify which characteristics such audits should meet according to users.

Javier Vijande, PhD Laura Oliver Cañamás
Therapy Physics 0 people interested