ESTRO 2024 - Abstract Book

S600

Clinical - Breast

ESTRO 2024

Mashhad, Iran, Islamic Republic of. 5 Iran University of Medical Sciences, Medical physics, Tehran, Iran, Islamic Republic of

Purpose/Objective:

The aim of this study was to investigate the advantages of dosiomics features over traditional dose-volume histogram (DVH) parameters in predicting skin toxicity after radiotherapy in breast cancer patients.(1-3)

Material/Methods:

This prospective study, included 76 breast cancer patients who underwent 3D conformal radiation therapy with a prescribed dose of 50 Gy in 25 fractions and 10 Gy as a boost in 5 fractions. Radiation toxicity was scored one week after the end of radiation therapy using Common Terminology Criteria for Adverse Events (CTCAE) version 5 that divided skin toxicity into four categories. A total of 140 parameters were extracted. 107 dose-segmented dosiomics features, including shape, first-order, and texture features derived from the dose distribution. Additionally, 33 clinical and dosimetric parameters were incorporated. The patients were randomly divided into two groups: one for model training (80%) and the other for test datasets (20%). To ensure uniformity across variables, the dosiomic features were standardized into z-scores with a mean of zero and a standard deviation (SD) of one. The feature dimension was reduced using the following methods: inter-feature correlation, which relied on Spearman's correlation coefficients, and feature importance, determined through a minimum Redundancy - maximum Relevance (mRmR) feature selection function.

These features were then evaluated using two machine-learning algorithms: the Support Vector Machine (SVM) and the Random Forest (RF).

Two different models were developed: (a) a model with DVH and clinical parameters (DVH + clinical model) and (b) a model with the selected dosiomic features and clinical parameters (dosiomic + clinical model). Nested sampling and hyper-tuning methods were adopted to train and validate the prediction models.

The performance of each algorithm was assessed in terms of accuracy, ROC curve, and AUC. (Figure 1)