ESTRO 2024 - Abstract Book

S3498

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

ESTRO 2024

1 Lund University, Department of Medical Radiation Physics, Lund, Sweden. 2 Skåne University Hospital, Department of Haematology, Oncology, and Radiation Physics, Lund, Sweden. 3 Lund University, Department of Translational Sciences, Medical Radiation Physics, Malmö, Sweden

Purpose/Objective:

The irradiated rectum volumes during prostate radiotherapy are largely dependent on the proximity or even overlap, between the two organs. Parts of the rectal wall can receive the full prescribed dose. Dose-volume constraints, based on conventional techniques and fractionation, aim to limit the risk of grade 2 and 3 late rectal toxicity [1]. A more recent review suggests that dose-volume constraints for grade 1 toxicity can be achieved with modern, intensity-modulated techniques [2]. To achieve the best possible trade-off between the prostate and rectum, timely optimization is required. In clinical routine, the achievable quality can be restricted by the time available for optimization before the scheduled treatment start. The purpose of this study was to evaluate the possible reduction in rectum doses for ultra-hypofractionated prostate radiotherapy using a Pareto front-based automated approach and to use it as training data for machine learning. A second objective was to investigate the correlation between achievable plan quality and patient specific features i.e. prostate/rectum overlap. Nineteen prostate cancer patients, representing various prostate sizes, were selected from the HYPO-RT-PC trial [3]. Patient-specific features (PSFs), such as rectal volume, PTV volume, bladder volume, whole-body structure volume, distances between PTV and body contour in four directions, distance from femoral heads to PTV, overlap between bladder and PTV, and overlap between rectum and PTV, were extracted. Automated treatment plans were generated using in-house software, resulting in 800 VMAT plans for each patient with varying dose-volume objectives. The identification of the Pareto front, consisting of clinically acceptable and deliverable treatment plans, was performed. Two machine learning algorithms, k-nearest neighbor (KNN) and deep autoencoder (DA), were utilized and trained on patient-specific features and CT image data, respectively. These algorithms aimed to replicate Pareto-optimal plans, which were then compared to assess their clinical feasibility. Plans for all patients, from different regions of the Pareto front —representing different priorities between the target coverage and the organ doses — were grouped and compared to the manually generated treatment plans. Data normality was evaluated using the Shapiro-Wilk test. A comparison between each Pareto level group and manual plans was made with a paired t-test (α = 0.005, Bonferroni corrected for multiple tests). PTV V95%, which had non-normally distributed data, was analyzed using a Wilcoxon signed-rank test. Material/Methods:

Differences between Pareto fronts were assessed with a 2D Kolmogorov-Smirnov test [4]. The significance of the linear regression model's slope (dose vs. rectum overlap) was tested using an F-test at α = 0.05.

Results:

Filtering of non-clinical and non-Pareto optimal plans resulted in a subset of 10 to 20 plans, representing 1.25 2.5% of the initial 800. An observable trend of lower rectal doses across the Pareto-optimal plans compared to manual plans was identified. For clinical dose-volume criteria, there was no significant difference between the automated plans with prioritized PTV coverage compared to the manually optimized plans. Automated plans consistently displayed smaller volumes receiving specified doses (rectal V90%, V75%, and V65%) compared to the manual plans. Surprisingly, maximum doses to the femoral heads were significantly lower in the automated plans. The machine learning models successfully recreated Pareto-optimal plans, demonstrating no significant