S2850
RTT - RTT education, training, and advanced practice
ESTRO 2026
Digital Poster 1247 integrating AI-based clinical assistants into radiation therapy practice: from CT simulation to
number 5289 collected data from localized prostate cancer patients treated with SBRT using MRgART (Mridian, ViewRay) in 5-fraction schemes, with 100% of sessions being adaptive. Treatment preparation involved the intake of one or two glasses of water 10- 15 minutes prior and an astringent diet before treatment. Bladder and rectum volumes at each session were recorded and compared to the simulation values, assessing reference constraints (Bladder V5<40Gy, Bladder V20<40Gy, Rectum V10<36Gy) and PTV coverage (V>95%). Results:
photon and proton treatment Han Min Tan 1 , Eric Pei Ping Pang 1,2
1 Division of Radiation Oncology, National Cancer Centre Singapore, Singapore, Singapore. 2 Oncology Academic Clinical Programme, Duke-NUS Graduate Medical School, Singapore, Singapore Purpose/Objective: Radiotherapy services typically maintain extensive clinical standard operating procedures (SOPs) to support staff training, standardisation of practice, and guidance for both routine and specialised procedures. Retrieving procedure specific details can be slow when users do not know the exact document title. We implemented a large language model (LLM)–based, SOP grounded chatbot to let radiation therapists (RTTs) obtain guidance through natural language queries, without complex search syntax. Material/Methods: Pair Assistant, an open government product (GovTech, Singapore), was deployed on the Government Commercial Cloud for public service workers and leveraged Claude 4 Sonnet with custom long-context processing. Prompt instructions were engineered to define the chatbot’s role within the radiation therapy domain, link the relevant work-instruction documents, and to utilise guardrails to minimise misinterpretation of common abbreviations and reduce hallucinations. Reference SOPs were reformatted into markdown with clear headings to facilitate instruction-based citation.We created sub-domain specific chatbots dedicated to CT simulation and treatment workflows (photon and proton), thereby improving relevance by narrowing the knowledge scope for each assistant. The chatbots underwent iterative testing using queries of increasing complexity; based on observed responses, both prompt instructions and reference documents were refined to enhance accuracy and consistency of domain specific answers. Results: Customised Pair Assistant chatbots were able to address CT simulation and treatment-related procedural queries with high precision. The assistants demonstrated the capability to request additional contextual details before providing an answer. An infographic summarising the development and deployment workflow is presented in Figure 1a, and utilisation metrics are shown in Figure 1b.
A total of 460 fractions from 92 patients were analyzed. Dose schemes were 36.25 Gy (50 patients) and 40 Gy (42 patients). The means for each session regarding OARs and PTV coverage are shown in Fig 1.A significant decrease in mean bladder volume was observed between simulation (152.84 cc) and treatment sessions (range: 104.19 cc - 117.54 cc). Rectal volume also showed variability (68.87 cc at simulation vs. 55.3 - 61.7 cc at treatment).Notably, PTV coverage (V95%) consistently improved, increasing from 92.6% in the simulation plan to an average exceeding 93.8% in the mean of the adapted fractions.Despite marked anatomical variability in each session, the adaptive strategy ensured 100% compliance with all dosimetric constraints in all fractions. Conclusion: Daily adaptation effectively compensates for inter- fractional anatomical variations, not only ensuring bladder and rectal protection but also optimizing PTV coverage. This brings into question whether it is only necessary to certify or consider bladder morphology, which can influence constraint coverage.These preliminary findings suggest the possibility of simplifying the clinical workflow and improving the patient experience without compromising treatment safety or efficacy. However, further studies with robust statistical analysis are required to confirm these results and establish stronger clinical recommendations. Keywords: Prostate, MRgRT, RTT
Made with FlippingBook - Share PDF online