¹û¶³´«Ã½

Before advanced image-processing techniques can be implemented clinically, their impact on treatment decisions must be evaluated. One example is their potential influence on daily treatment plan selection, which is currently based on CBCT images. demonstrates with her research that AI-based image quality improvements, also known as synthetic CT (sCT) generation, reduce inter-observer variability in treatment plan selection. However, to fully benefit from more consistent plan selection due to enhanced image quality, clear guidelines must first be established.

Auto-segmentation and automatic plan selection

In addition to better image quality, auto-segmentation can also be beneficial. This technique enables a computer program to automatically recognize and delineate organs on CBCT images. To further tailor the treatment plan to daily anatomy, robust auto-segmentation of daily CBCT is essential. However, CBCT images are often not manually contoured due to their low quality and the time-consuming nature of the task. Yvonne de Hond investigates within her research whether an AI-based auto-segmentation model for CBCT images can be trained using adapted CT images. The results show that this is feasible, meaning these automatic contours can be used to improve the adaptation of radiation therapy to daily anatomy. Additionally, Yvonne de Hond examines whether automatic selection of the treatment plan performs as well as manual selection. The findings indicate no significant differences in delivered dose, suggesting that automatic plan selection can enable faster and more consistent identification of the optimal treatment plan each day.

Online adaptive re-planning

To further reduce radiation dose to healthy organs, a more advanced technique is required: online adaptive re-planning. In this approach, the treatment plan is adjusted per treatment based on the daily anatomy. Synthetic CT images play a crucial role in this process, as CBCT images are not accurate enough for the required dose calculations. This research evaluates the quality of sCT images, not only in terms of image quality but also in their anatomical accuracy. This is essential to ensure that treatment planning remains reliable.

Improving online adaptive radiotherapy

A key insight from this research is that standard methods for assessing image quality are not always sufficient. Some AI-models generate images that appear visually accurate but fail to represent the patient’s anatomy correctly. Therefore, a method to detect and visualize these inaccuracies was developed, helping clinicians make decisions based only on reliable image regions. The research contributes to improving and making online adaptive radiotherapy on C-arm Linacs more accessible. This allows more patients to benefit from reduced side effects due to lower radiation exposure to surrounding organs while maintaining adequate tumor coverage. However, further research is needed to automate the process and reduce treatment duration, enabling more patients to receive adaptive radiotherapy. Additionally, future studies should focus on the safe implementation and quality assurance (QA) of AI in clinical practice.

This research was done in collaboration with Catharina Ziekenhuis Eindhoven and Elekta Ltd. (Crawley, UK).

Title of PhD-thesis: . Supervisors: Prof. Coen Hurkmans, Dr. Paul van Haaren and Dr. Rob Tijssen.