ESTRO 2024 - Abstract Book

S4455

Physics - Machine learning models and clinical applications

ESTRO 2024

Two ML models were developed to early identify prostates with pelvic nodes plans that could fail PSQA. The models were tuned to be particularly sensitive, being sure to identify low score plans, but allowing to reduce the PSQA workload. XGBoost shows better performances compared to Random Forest, but the tendency of overfitting. We found good results building models on simple explainable metrics and we underline the importance of having a homogenous population to build the models on. The models validation will continue as increasing the number of patients considered. Moreover, similar models for other anatomical regions/treatment types will be developed.

Keywords: PSQA, ML models , Complexity metrics

References:

1. Younge et al. (2016) "Predicting deliverability of volumetric-modulated arc therapy (VMAT) plans using aperture complexity analysis";

2. Du et al. (2014) "Quantification of beam complexity in intensity-modulated radiation therapy treatment plans";

3. Chiavassa et al. (2019) "Complexity metrics for IMRT and VMAT plans: a review of current literature and applications";

4. Han et al. (2023) "Integrating plan complexity and dosiomics features with deep learning in patient-specific quality assurance for volumetric modulated arc therapy";

1124

Digital Poster

Feasibility of AI chosen automated couch and gantry beam angles for brain tumors treated with IMPT

Robert Kaderka 1 , Yan-Cheng Huang 2 , Hsin-Chih Lo 2 , Ethan Tu 2 , Che Lin 2,3 , Chang Chang 4,5

1 University of Miami Miller School of Medicine, Radiation Oncology, Miami, USA. 2 Taiwan AI Labs, Taiwan AI Labs, Taipei, Taiwan. 3 National Taiwan University, Department of Electrical Engineering, Taipei, Taiwan. 4 University of California San Diego, Radiation Medicine and Applied Sciences, La Jolla, USA. 5 California Protons Cancer Therapy Center, Radiation Oncology, San Diego, USA

Purpose/Objective:

Compared to standard photon therapy, proton therapy can reduce the dose to normal tissue but requires more planning resources. Achievable plan quality in proton therapy is highly dependent on the chosen beam arrangement. In a previous study, an AI tool was developed and demonstrated to automatically create coplanar beam angles for targets in the liver 1 . However, when targeting the brain, more complex beam arrangements with non-coplanar fields are often used. Creating these arrangements takes significant time and requires a high level of skill and experience from the human planner. This study explores AI's potential to automatically select couch and gantry angles for non-