ESTRO 2025 - Abstract Book

S3395

Physics - Machine learning models and clinical applications

ESTRO 2025

Conclusion: These findings suggest predictive FMCIB features are impacted only when non-GTV voxels are disturbed, implying the model disregards some or all GTV data. Further investigation into foundation model feature interpretation is warranted.

Keywords: Foundation model, deep learning, quality control

References: 1 M. Tortora et al., “Radiomics applications in head and neck tumor imaging: A narrative review,” Cancers (Basel), vol. 15, no. 4, p. 1174, Feb. 2023. 2 M. L. Welch et al., “Vulnerabilities of radiomic signature development: The need for safeguards,” Radiother. Oncol., vol. 130, pp. 2 – 9, Jan. 2019. 3 S. Pai et al., “Foundation model for cancer imaging biomarkers,” Nat. Mach. Intell., vol. 6, no. 3, pp. 354– 367, Mar. 2024. 4 M. L. Welch et al., “RADCURE Version 3,” The Cancer Imaging Archive, 2023. doi: 10.7937/J47W -NM11. 5 L. Wee et al., “HEAD -NECK-RADIOMICS- HN1,” The Cancer Imaging Archive, 2019. doi: 10.7937/tcia.2019.8kap372n. 6 A. Grossberg et al., “HNSCC Version 4,” The Cancer Imaging Archive, 2022. doi: 10.7937/k9/tcia.2020.a8sh -7363.

2263

Digital Poster A clinical decision-making assistant for streamlining prostate MR-guided adaptive radiotherapy Fanchi Su, Siqiu Wang, Brien Washington, Ti Bai, Daniel Yang, Aurelie Garant, Raquibul Hannan, Neil Desai, Mu-Han Lin Radiation Oncology, UTSW, Dallas, USA Purpose/Objective: Our previous studies established the feasibility of a deep-learning-based dose predictor for accurate adapt-to-shape (ATS) dose estimation and conducted extensive dosimetric analyses of 100 fractions of prostate stereotactic ablative radiotherapy (SAbR). While the ATS adaptation workflow can enhance or maintain pre-plan quality by fully addressing daily anatomical changes, results demonstrated that adapt-to-position (ATP) adaptation achieved reasonably good plan quality in one-third of cases and significantly improved workflow efficiency, resource utilization, and adaptation time. Strong correlations were identified between anatomical volume changes and discrepancies in maximum doses at bladder-target and rectum-target interfaces. These findings suggest that a deterministic dose model, based on dose differences between ATS and pre-plan, can effectively assess ATP dosimetry. This study advances the deep-learning-based dose predictor by incorporating key dosimetric indicators to support decision-making between ATS and ATP workflows in prostate MR-guided adaptive radiotherapy (MRgART). Material/Methods: Figure 1 illustrates the workflow of the proposed decision-making assistant. Daily MR images serve as the initial input. AI-based contouring is applied to these images to account for daily deformations, anatomical fillings, and interfractional organ changes. The ATS dosimetric predictor, a deep-learning-based dose prediction model, estimates PTV coverage and OAR doses. Simultaneously, pre-plan dose information and plans are analyzed to determine if ATP is recommended. This assessment evaluates changes in maximal doses (Dmax) to the bladder and rectal walls, as well as differences in overlapped volumes (OV) at the bladder-target and rectum-target interfaces, comparing the pre-plan with ATS dose estimates.