Prospective Evaluation of AI-Based BiCycle Autoplanning for Advanced Cervical Cancer Brachytherapy
Linda Rossi, Rik Bijman, Henrike Westerveld, Michèle Huge, Inger-Karine Kolkman-Deurloo, Miranda Christianen, Lorne Luthart, Jan Willem Mens, Huda Abusaris, Raymond de Boer, Sebastiaan Breedveld, Ben Heijmen, Remi Nout.
Erasmus MC Cancer Institute, Department of Radiation Oncology, Rotterdam, The Netherlands. Radiotherapy & Oncology, 2025, doi:10.1016/j.radonc.2025.111029
What was your motivation for initiating this study?
Brachytherapy is an essential part of treatment for locally advanced cervical cancer, but manual planning is labour-intensive and operator-dependent. Our previous study clearly indicated the need to accelerate plan preparation in order to reduce patient discomfort, pain and anxiety [1].
At Erasmus MC, we have developed an in-house, rule-driven artificial intelligence (AI) autoplanning system, BiCycle, in which dwell-times are automatically optimised to fulfil a list of rules that we call a “wish-list”. Retrospective studies indicated that BiCycle could generate treatment plans of equal or higher quality compared with manual ones, in both intracavitary and interstitial cases [2]. Dosimetric dose distribution as well as loading pattern (i.e. smoothed dwell positions, desired channel contribution ratio, etc.) could be controlled and optimised with BiCycle.
However, retrospective comparisons alone cannot establish clinical feasibility or, most importantly, planning time in real contexts. Our motivation, therefore, was to evaluate BiCycle prospectively in a real clinical workflow, testing whether AI-generated plans could maintain or improve plan quality and achieve clinical approval while significantly reducing planning time.
What were the main challenges during the work
Conducting a prospective study within an active clinical environment presented both technical and organisational challenges. The primary goal was to integrate AI-based autoplanning alongside routine clinical operations without disrupting the workflow.
The first challenge was to design a study protocol that was feasible within clinical schedules while maintaining scientific and clinical relevance. Ensuring clinical acceptance and oversight represented another key difficulty. While the use of AI can automate optimisation, the treating physicians must ultimately approve or be able to adjust each plan. It was essential to establish a process for review and modification of AI-generated plans (Auto_Adj) in precisely the same way as is done routinely in clinical practice (i.e. utilising the same software, tools, and setup) in order to maintain clinical responsibility and relevant evaluation.
Time constraints posed an additional limitation. Three out of 41 plans were excluded because they were not evaluated within the predefined time window, the imposition of which was a necessary restriction to reflect real-world conditions. Similarly, unlike retrospective studies, prospective data collection allowed no opportunity to recover missing information. Planning times were inadvertently not recorded for five of the remaining 37 plans, so these cases had incomplete data.
All these aspects required meticulous preparation before the study began, so the whole design was tested with five pilot patients who were not included in the analysis. When patient recruitment started, meticulous coordination and collaboration between radiation oncologists, radiotherapy technologists, and medical physicists were essential.
What are the most important findings of your study?
The study demonstrated that AI-based autoplanning can match or surpass the quality of manually optimised plans while substantially improving efficiency.
Across all 37 fractions that were analysed, the AI-generated (Auto_Adj) plans were rated equivalent or superior by treating physicians in 95% of cases, and 76% were deemed superior. Target coverage for the high-risk and residual gross tumour volumes (CTVHR, GTVRES) remained comparable with those of manual plans, with only a marginal reduction in intermediate-risk target coverage (CTVIR D98%). The latter was considered acceptable by physicians, as indicated by the superior target score of the automated plans.
Importantly, the use of BiCycle led to significantly better sparing of organs-at-risk; the total equivalent dose in 2Gy fractions delivered to a volume of 2cm3 of healthy tissues (EQD2 D₂cm³) was reduced for bladder by 3.7Gy, rectum by 3.0Gy, sigmoid by 1.0Gy, and bowel by 1.4Gy relative to the clinically delivered plans.
The most striking operational result was the reduction in total planning time — from an average of 44.1 minutes for manual planning to 9.4 minutes using BiCycle autoplanning (including minor physician adjustments). In terms of actual hands-on time, i.e. user engagement with the treatment planning system, the reduction went from 44.1 minutes for manual planning to 4.0 minutes for automated. This significant decrease highlights the transformative potential of AI to streamline brachytherapy planning workflows.
What are the implications of this research?
This study provides the first prospective validation of an AI-based autoplanning system in cervical cancer brachytherapy. It has demonstrated that AI-based solutions are both feasible and beneficial in daily practice. The use of autoplanning can significantly reduce treatment planning time and meet physicians’ approval, without compromising safety.
Following these findings, the BiCycle system has been implemented clinically after regulatory clearance under the EU medical device regulation and is used daily in our clinic. The success of BiCycle establishes a framework for further AI applications in the preparation of plans for cervical brachytherapy and across other brachytherapy sites.
This study’s results also emphasise the necessity for further advancements in the brachytherapy workflow, in which AI can help to alleviate workload while enhancing or preserving quality. Such advancements can benefit both healthcare providers and patients, and increased efforts from institutions and vendors should be directed toward this goal.
Dr Linda Rossi
Medical physicist
Department of Radiation Oncology
Erasmus Medical Centre (Erasmus MC)
Rotterdam, The Netherlands.
l.rossi@erasmusmc.nl
References
[1] van Vliet-Perez S, van Paassen R, Wauben L, Straathof R, Berg N, Dankelman J, Heijmen B, Kolkman-Deurloo IK, Nout R. Time action and patient experience analyses of locally advanced cervical cancer brachytherapy. Brachytherapy 23 (2024) 274-281.
[2] Rossi L, Bijman R, Chopra S, Mittal P, Panda S, Westerveld H, Christianen M, Kolkman-Deurloo IK, Breedveld S, Nout R, Heijmen B. Rule-based AI automated adaptive treatment planning for image-guided cervical cancer brachytherapy. Brachytherapy 24 (2025) 711–720