Dosimetric comparison of manual and automatic AI-assisted registration for assessment of interfraction variations in MR-guided adaptive brachytherapy.

Abstract

The standard treatment for locally advanced cervical cancer is external beam radiotherapy followed by a brachytherapy boost. This method leads to excellent clinical results. At the Department of Radiation Oncology of the Vienna General Hospital/ MedUni Vienna, such a boost is typically administered in four fractions of high-dose rate (HDR) brachytherapy. From a clinical point of view, organ motion between fractions has been shown to be within acceptable ranges for most patients. Therefore, the same treatment plan can be used for two consecutive days of treatment. However, some patients present with significant organ motion. These cases require intervention of the planning team. The process of assessing interfraction variations is currently performed either visually or by landmark-based rigid image registration, which is a manual and time-consuming process, and therefore difficult to implement in clinical practice. Currently there are no solutions for rapid quantitative evaluation of organ at risk interfraction motion in commercial treatment planning systems (TPS). Novel deep learning methods based on convolutional neural networks have shown great promise to overcome this challenge. A detailed comparison between the dosimetric impact of both methods is still pending. The aim of this thesis is to quantitatively assess the difference between standard manual TPS-based applicator registration and novel artificial intelligence (AI) - based method by analyzing the dosimetric impact of registration uncertainties for both methods. To achieve this goal, a test cohort of 9 female cervical cancer patients (N=9) underwent a repetitive imaging protocol during treatment. These image sets were registered for each subject using both the conventional manual method (which served as a benchmark) and the algorithm supported by AI. The impact of registration uncertainties was evaluated via discrete dose volume histogram parameters of organs at risk. The results of this study show that there are differences in how organs at risk are affected by existing registration uncertainties of the AI-based registration method. For bladder and sigmoid, the dosimetric impact of the registration uncertainties is more pronounced than for bowel and rectum. The analysis of inter- and intrafraction motion highlighted the importance of integrating a fully automated assessment workflow in the future. This study underscores the promising role of AI in enhancing brachytherapy treatment planning, suggesting further research into its integration could significantly improve clinical outcomes for patients with cervical cancer and reduce the workload for the cervical cancer treatment.