ESTRO 2025 - Abstract Book

S3238

Physics - Intra-fraction motion management and real-time adaptive radiotherapy

ESTRO 2025

Conclusion: This work provides an initial evidence basis for MR-guided bladder radiotherapy specific CTV to PTV population margins. Personalised margins are feasible and can be generated rapidly using the workflow created. Future work will perform a dosimetric analysis comparing original, updated and personalised margins.

Keywords: margins, personalisation, MR-guided

References: [1] Kim, H. S. et al. (2012). Bidirectional local distance measure for comparing segmentations. Medical Physics, 39(11), 6779-6790. https://doi.org/10.1118/1.4754802 [2] Van Herk et al. (2000). The probability of correct target dosage: dose-population histograms for deriving treatment margins in radiotherapy . International Journal of Radiation Oncology, Biology, Physics, 47(4), 1121–1135. https://doi.org/10.1016/S0360-3016(00)00518-6

1896

Digital Poster Explainable AI for deep learning-based markerless lung tumor tracking Dragos Grama 1,2 , Max Dahele 2 , Ben Slotman 2 , Wilko F.A.R. Verbakel 2,3 1 R&D, Synaptiq, Cluj Napoca, Romania. 2 Radiation Oncology, Amsterdam UMC, Amsterdam, Netherlands. 3 R&D, Varian Medical Systems, Palo Alto, USA Purpose/Objective: Tumor position monitoring using fluoroscopic (kV) images acquired during Volumetric Modulated Arc Therapy (VMAT) delivery has become available in a research setting. Accurate lung tumor tracking during stereotactic body radiotherapy (SBRT), using VMAT, can help to ensure that the tumor is only irradiated when it is inside the planning target volume 1 . Traditionally, template matching is used to determine the tumor position, but with low accuracy and tracking rates. A deep learning-based approach has the potential to improve this, but as deep learning is considered a "black box", it would be desirable to know when to trust the predictions made by the model. In this work, we investigate the feasibility of explainable AI (XAI) methods for deep learning-based markerless lung tumor tracking, to highlight relevant features for AI-based tracking models. Such information could be used e.g. during treatment delivery to know when not to trust the tracking algorithm and pause the radiation delivery. Material/Methods: We conduct several experiments on a Quasar phantom that includes a high density spherical tumor and six clinical patients with small lung tumors (0.23 - 2.93 cm 3 ), during delivery of typical SBRT. The suitability of the XAI methods for tumor tracking is assessed on both phantom and patient data, by visually inspecting the generated heatmaps according to each XAI method. Four XAI methods are evaluated, i.e. Guided Backpropagation 2 (GBP), Layer-wise Relevance Propagation 3 (LRP), DeepLIFT 4 and PatternAttribtuion 5 , for highlighting the most relevant features to a deep Siamese markerless lung tumor tracking model 6 . Results: For the phantom, GBP, LRP and DeepLIFT highlight the boundary of the tumor as being important for determining its position (Figure 1), which is the expected behavior for tracking a high density sphere surrounded by lower density tissue.