ESTRO 2024 - Abstract Book

S3529

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

Both models were trained individually on a dataset of 513 VMAT prostate patients and validated on 57 patients. The fully connected pipeline was then tested with a new data set of 21 patients. The prescribed dose was 60 Gy to the PTV in 20 fractions. To evaluate the performance of this pipeline, the treatment plans were recalculated and reviewed. Each plan was quantitatively scored based on clinically established clinical goals for both the target and the OARs, in order to ascertain their clinical acceptability.

Results:

Integrating both stages for the first time, our pipeline's output reflected a commendable plan quality. The generated plans could meet (83.7±8.8) % of the clinical goals. The cases in which the clinical goals were not met, were typically associated with the maximum dose, where the allowed max dose was exceeded by not more than 5%, or the coverage V95% of the PTV, where our models achieved slightly lower target values of up to 4% in some cases. The total time to produce a deliverable treatment plan was on average (39±5) seconds and required a single command to execute. The majority of this time was spent on the preprocessing (26 seconds) of the CT and including structure segmentations. The dose prediction and plan generation each took about 6 seconds.

Conclusion:

To our knowledge, we created the fastest pipeline for fully automated treatment planning. Additionally, our pipeline is capable of generating a deliverable treatment plan outside a TPS. In this pilot study, we connected both of our previously developed models for the first time and created a fully automated workflow, combined with a commercial solution for auto-segmentation. Although the pipeline occasionally did not meet all clinical objectives for an acceptable treatment plan, this innovative approach demonstrates its potential. Current work in progress aims to further improve this model and embed it into a clinical workflow with automatic data transfer between the systems and the plan review being performed on a dedicated platform (e.g. a secondary dose check software). Our architecture is much faster and more efficient than existing approaches, making it a promising candidate for future real-time or online adaptive radiotherapy.

Keywords: deep learning, automation, treatment planning

References:

Heilemann, G., Zimmermann, L., Schotola, R., Lechner, W., Peer, M., Widder, J., Goldner, G., Georg, D., & Kuess, P. (2023). Generating deliverable DICOM RT treatment plans for prostate VMAT by predicting MLC motion sequences with an encoder ‐ decoder network. Med Phys, 50(8), 5088–5094. https://doi.org/10.1002/mp.16545 Zimmermann, L., Faustmann, E., Ramsl, C., Georg, D., & Heilemann, G. (2021). Technical Note: Dose prediction for radiation therapy using feature-based losses and One Cycle Learning. Medical Physics, 48(9), 5562–5566. https://doi.org/10.1002/mp.14774

1014

Digital Poster

Dosimetric evaluation of accelerated partial breast irradiation using VMAT technique