ESTRO 2024 - Abstract Book

S4418

Physics - Machine learning models and clinical applications

ESTRO 2024

3. Bauwens (2021) Radiotherapy and Oncology 157, 122-129

390

Digital Poster

Personalization of PTV margins based on SGRT and Big Data tools

Mathieu Gonod 1 , Laurent Delcoudert 1 , Stéphane Morisset 2 , Victor Robineau 1 , Jad Farah 3 , Léone Aubignac 1 , Igor Bessières 1 1 Centre G.F. Leclerc, Medical Physics, Dijon, France. 2 Consultant, Statistics, Perouges, France. 3 VisionRT, Sales and Clinical applications, London, United Kingdom

Purpose/Objective:

Despite the tremendous efforts in performing individualized radiation therapy, standard PTV margins are still commonly applied based on the treatment technique and tumor localization. However, as described in ICRU 62, PTV margins are added to account for different uncertainty components such as CTV movement, technical/delivery uncertainties and most importantly intra-fraction patient motion. The latter is typically tracked and recorded by Surface guided radiation therapy (SGRT) systems but this data remains under-exploited. The present work aimed at using big data tools to correlate intra-fraction motion to various clinical and technical parameters in order to define motion groups and identify opportunities for individualized and optimized PTV margins.

Material/Methods:

The motion amplitude of 5599 treatment fractions corresponding to 379 patients were extracted from AlignRT treatment reports (Vision RT Ltd., London, UK) using an in-house Python script. A second script was used to collect, from the hospital’s NIS, a total of 50 clinical/technical parameters such as treatment site, age, gender, immobilization device, dose prescription, etc. . The statistical distribution of intra-fraction motion amplitude was next analyzed. For patients with identical PTV margin, Conditional Inference Trees (CIT) were firstly used to generate motion groups with significant correlation between clinical/technical data and the maximum magnitude value of intra-fraction motion registered by AlignRT. Assuming that the PTV margin is defined so that the CTV is always within the 95% isodose, potential PTV individualization and margin optimization was calculated. Lastly, the predictivity of the CIT-based motion groups was evaluated using randomly selected test data representing 25% of the sessions and the cross validation was repeated four times to test its robustness on all data sets.

Results:

For PTV margins of 10 mm, 7-8 mm and 5 mm, the CIT algorithm helped identifying 10, 5 and 2 different motion groups respectively.