ESTRO 2024 - Abstract Book

S4535

Physics - Machine learning models and clinical applications

ESTRO 2024

[1] C. Sini, Dose–volume effects for pelvic bone marrow in predicting hematological toxicity in prostate cancer radiotherapy with pelvic node irradiation, Radiotherapy and Oncology, Volume 118, Issue 1, 2016, Pages 79-84, ISSN 0167-8140, https://doi.org/10.1016/j.radonc.2015.11.020. [2] Mell L.K. et al. Association between bone marrow dosimetric parameters and acute hematologic toxicity in anal cancer patients treated with concurrent chemotherapy and intensity-modulated radiotherapy. International Journal of Radiation Oncology Biology Physics 2008; 70:1431-1437, https://doi.org/10.1016/j.ijrobp.2007.08.074.

2454

Digital Poster

Can AI help to identify beams prone to fail pre-treatment verification?

Helena Vivancos Bargalló, Pedro Gallego, Natalia Tejedor, Cristina Ansón, Pablo Carrasco, Jaime Pérez-Alija, Fátima Leo, Marta Barceló, Agustí Ruiz, Alejandro Domínguez, Víctor Riu, Núria Jornet

Hospital de la Santa Creu i Sant Pau, Medical Physics, Barcelona, Spain

Purpose/Objective:

VMAT and IMRT treatments represent a 70% of total treatments in our centre. Pre-treatment verifications remain common to assure that treatment unit can deliver the planned dose fluence. Our study aims to train a classifier to predict pre-treatment verification failures, saving LINAC measurement time.

Material/Methods:

3196 VMAT beams and 3792 IMRT beams (Table 1) planned for four LINACs (three TrueBeams: TB0, TB1 and TB3 and one Clinac 2100, Varian) were used for this study. For each beam a set of complexity metrics (Table 1) were calculated. Pre-treatment verifications using Portal Dosimetry (Varian vs 15.6) were performed for each beam. A global analysis with a 10% low dose threshold and a passing/fail threshold of 98% was used. Complexity metrics and gamma analysis results (pass/fail) for TB0 (VMAT) and for TB0 and TB1 (IMRT) were used as training sets for 2 Random Forest Classifiers (one per technique). For hyperparameter tuning, a grid search employing 10-fold cross validation was performed in those training sets. For IMRT, we needed data from two Truebeams due to limited failed beams. This approach was adopted to avoid having a too limited dataset or overfitting to beams that passed the criteria. First, we tested the classifiers on a different dataset from the Truebeams used for training. Additionally, the models were tested on datasets from two different treatment units, TB3 and Clinac (Table 1). Random Forest classifiers have the advantage of providing information on the importance of each feature in the classification.

Table 1: Summary of beams used for training/testing the models

Training/Cross Validation beams

Technique

Pathologies

Complexity metrics

Test beams