ESTRO 2024 - Abstract Book

S4560

Physics - Machine learning models and clinical applications

ESTRO 2024

2743

Digital Poster

Machine learning-aided automatable replanning tool to enhance deliverability in modulated RT plans

Caterina Zaccone 1,2 , Nicola Lambri 2,3 , Monica Bianchi 1,4 , Andrea Bresolin 4 , Simone Buzzi 1,4 , Damiano Dei 4 , Pasqualina Gallo 4 , Francesco La Fauci 2 , Francesca Lobefalo 2 , Lucia Paganini 2 , Sara Parabicoli 2,1 , Marco Pelizzoli 2,1 , Giacomo Reggiori 2 , Stefano Tomatis 2 , Marta Scorsetti 2,3 , Cristina Lenardi 1 , Pietro Mancosu 2 1 Università degli studi di Milano, Physics Department, Milan, Italy. 2 IRCCS Humanitas Research Hospital, Radiotherapy and Radiosurgery Department, Rozzano, Milan, Italy. 3 Humanitas University, Biomedical Sciences department, Milan, Italy. 4 IRCCS Humanitas Research Hospital, Radiotherapy and Radiosurgery Department, Rozzano, Milano, Italy

Purpose/Objective:

In the realm of radiotherapy, the challenge of over-modulation in treatment plans can substantially hinder their deliverability. The evaluation of gamma passing rate (GPR) under specific criteria - e.g., GPR(3%, 1 mm) < 90-95% - is a standard indicator for identifying plans with suboptimal deliverability. Recently, in our centre a machine learning (ML) model was trained on a large dataset of complexity metrics from over 15,000 arcs, aiming to predict the GPR using a stringent 3%/1 mm criteria with a 10% threshold [1]. The aim of this study was to investigate whether an automatable replanning process, supported by our ML model, can effectively enhance the deliverability of patients, considering different region of treatment, whose predicted GPR falls significantly below 95%, while maintaining high dosimetric quality.

Material/Methods:

The ML model was tested on the latest 2,507 patients - comprising over 5,900 treatment arcs - who underwent treatment with volumetric modulated arc therapy (VMAT) using Eclipse15 (Varian Medical Systems, Palo Alto, CA) in our center during 2022-2023. From the cohort of the 100 patients with the lowest mean GPR predictions, ten patients from different treatment region were chosen. These plans were re-optimized from scratch using the same dose volume objectives as the original plans. Two automatable approaches were considered: (i) limiting the Monitor Units to 70% of the total original count (MUlimit), (ii) utilizing the recently introduced Eclipse TPS Aperture Shape Controller (ASC) tool. The re-optimized plans were compared to the original ones by evaluating any potential improvement in terms of: (a) modulation complexity score (MCS) to assess plan complexity, (b) predicted_GPR(3%,1mm), (c) patient specific QA (PSQA) by GPR(3%,1mm), and (d) dose-volume histogram (DVH) points for specific constraints related to the specific treatment regions. In detail, for (d) the differences between the re-optimized plan and the original plan were evaluated aggregating the values for the Organs At Risks (OARs), Body and PTV/CTV (Planning/Clinical Target Volume) separately with the intention of employing this information for distinct treatment regions. The Wilcoxon signed rank test was used for statistical analysis (p<0.05 significant).

Results:

We analysed 102 arcs from 30 treatment plans of the 10 selected patients (see an example in Figure 1). The re optimized plans displayed a significant reduction in complexity, evidenced by a median increase in MCS from 0.17